1. See above, Magic's covered everything

2. If this is the typical circuit you're referring to, then it helps to understand how the controller is being used. The controller has an error amp inside, set to adjust VFB to 2.50V. It controls the CMP pin. The CMP pin in turn is read by the current comparator, this being the setpoint for where to turn off the switch, implementing peak current mode control.
So, in short, V(CMP) ~ Iout.
In a common-ground application, the error amp is used directly, and there is a voltage divider from output to VFB, and a compensation R+C from VFB to CMP. The compensation accounts for the phase shift of the converter and output filter cap, stabilizing the output.
A note about control: you'll often see descriptions of control systems in terms of "if-then" events, or "greater-lesser-than" digital comparisons. This is wrong: control is a continuous, linear process. (Well, it
can be done by comparison -- call hysteretic control. But this case happens to be the linear kind.) The aim here is to adjust the output, from moment to moment, gently enough that it doesn't run away and oscillate, while -- over the long term -- adjusting it so that the output is as close to the setpoint as possible.
Anyway, that's common ground; what if our output is isolated? We can use an optoisolator to communicate the output voltage to the controller. But these are
not accurate devices. We can put a resistor to the LED, driving it with a current proportional to the output voltage, and measure a current at the phototransistor that's -- who knows, maybe 50%, maybe 300%! It varies with part number, manufacture (gain is not necessarily consistent from part to part) and age (primarily the LED fading with use, over 1000s of hrs). So that's pretty useless.
So what we do instead, is use the controller as a transconductance amplifier. That is, it's making a current (our output), and we adjust that current by the input voltage. So it has gain in units of A/V (conductance), and the amps and volts are measured in different locations, hence "trans". Transconductance. We set that voltage with the opto, and set the opto's current* with a separate error amplifier -- this time referenced to the output supply, so we can get an accurate voltage at the output.
*Okay so the opto has current gain, converted to voltage with a resistor, and back to current again by the converter, so this is overall a current amplifier now. If it matters.

The key is, by placing the opto inside the control loop, yes it might be in error, but we are controlling that error, and as long as we've properly placed the min/max limits of the system (opto at minimum CTR (current gain) is still able to deliver full voltage to the controller; opto at maximum CTR does not cause the control loop to go unstable), we're golden -- the error is absorbed by the error amp.
And for that error amp, we have a convenient device, a TL431 usually (and many relatives). This is drawn as an adjustable zener diode, whatever the hell that's supposed to mean -- but the characteristic is nothing like a zener, in fact it is a three-terminal op-amp with an open-collector output and an inverting input pin with a massive (but conspicuously stable) offset voltage of 2.50V. Seen in this light, we don't need to have REF tied to "cathode" by a voltage divider (the standard "adjustable zener" configuration), but we can use a whole system to provide that feedback, and just like the error amp on the UC3842, we put an R+C across it ("cathode" to REF) to compensate its response.
So that's the complete system -- we can configure the UC3842 as a simple current amplifier by either strong-arming its CMP pin directly (which can't source/sink much current, so is easy to overdrive; in this case, VFB can be tied off to GND or something), or by wiring it as a unity gain inverting amplifier (say 10k from CMP to VFB, 10k from VFB to opto).
3. Not sure what you mean about reversal or biasing -- it's a class D circuit, switching, no bias needed. It's designed to spend as little time as possible in the linear range, to save on power dissipation.
The usual configuration is a low-side switch (source and gate reference to GND, give or take), with the primary in the drain circuit (from drain to +V). The peak voltage (when the MOSFET turns off) shoots above +V, driven by the transformer's inductance.
It sounds like there's an R+C damper (or snubber), which probably absorbs the free ringdown of the transformer's inductance with the MOSFET drain capacitance. This helps reduce lower frequency EMI, at the expense of some power dissipation (mostly in the resistor).
It sounds like there's also a peak clamp snubber: when the drain voltage shoots up (flyback), the diode clamps that voltage into the 33nF cap (which is referenced to +V, because the primary is). Actually once the diode is conducting, it's a resonant circuit of Lp (primary inductance) and 33nF, but this is so much lower (in impedance and resonant frequency) that it looks like a flat topped waveform, i.e., it seems to be clamping the voltage peak and that's it. The 100k then bleeds away the excess charge the capacitor has picked up; this is wasted, costing efficiency (there are configurations where this can be saved -- namely, the two-switch flyback or forward converter; these are more complicated to construct, though).
The clamping wouldn't be necessary if the transformer were ideal, but real transformers have leakage inductance as well. This appears as inductance in series between primary and secondary. In effect, the primary can't instantly tell what the secondary is doing, but rather that action is communicated through this inductance. As a result, when the secondary side diode turns on, most of the transformer inductance (charged as it was by the MOSFET) is clamped into the output filter cap (delivering useful power), but a fraction of that energy is trapped in the leakage inductance, and causes additional peak voltage on the primary side. This can destroy the transistor, or at least require a higher rated one, so it's better to clamp it instead. It's also wasted power -- there's no way to divert that leakage energy into the secondary, you're already doing all that you can on the secondary side.
So the best way to save this loss, is to optimize the transformer design so it isn't storing a lot of energy as leakage inductance.
4. Hah, nah, resistor dividers and pass regulators are 100% (or more) waste. Use multiple outputs -- say, a transformer with a 12V and 4V secondary, and then use an LDO to drop 4V down to 3.3V with good stability. Logic supplies can be picky, and cross-regulation (that is, the amount one supply's voltage changes, due to a change in load current on the other) tends to be mediocre, so it can be worthwhile overshooting just a bit and still using a lossy regulator but only for a volt or so (so the overall efficiency is still good).
If you can't get custom transformers made this way, you can stack transformers (at some cost to performance). For example, use "12V" and "3.3V" transformers, wired in parallel at the primary; the total load inductance is the parallel combination, so choose them appropriately for nominal load. Or, use a single "12V" transformer, then put a trans/autoformer in parallel with that secondary, to divide it down to "3V" or "4V". You can use one 1:2, or a pair of 1:1, coupled inductors in this way. There are also multi-winding configurable transformers out there, at some additional cost.
Or, at worst -- just make two independent converters. Can share the EMI filter and bulk (DC) cap. Can even buy off-the-shelf modules instead of building your own -- a great way to start if you're unfamiliar with safety regulations!
Cheers,
Tim