If frequency adjustment via variable inductor/capacitor is acceptable, this gets much easier. 200Vrms on 100pF at 3MHz is 75 VAR. With a Q of 20, you would only need 3.7W drive power, under 10W total if a class AB amplifier is used (and that's not really necessary, if a slightly distorted waveform (filtered square wave) is acceptable). Probably a Q over 100 would be typical.
To drive this, you would use an amplifier with variable output amplitude (either a class D amp with variable pulse width or variable supply voltage, or a class AB amp with variable signal level), and an adjustable LC matching network. The network both transforms the impedance to a suitable range (the amplifier requires a load of such-and-such impedance), and causes the capacitance to resonate at the desired frequency. A voltage or current sensor provides feedback to the amplifier, allowing it to oscillate (for the linear amplifier, a slow, nonlinear ballast is necessary, classically a light bulb, but an analog multiplier could also be used; for the switching amplifier, a comparator is all that's needed). Alternately, a PLL is used, which oscillates only within the desired frequency range, and tracks output phase (this is preferred for a wide amplitude range).
The coupling transformer design should be made carefully, to maintain balanced operation. Typically, a single ended primary winding would be placed, in a single layer, followed by a slitted cylinder foil shield, and then the secondary windings, bifilar, single layer, connected in series to obtain a CT winding. This is easiest on a cylindrical bobbin (various core shapes: E, U, P, etc.), but a toroid can be used.
The transformer should probably be in parallel with the quadrupole, to preserve that balance. This requires a pretty high impedance transformer (100pF at 3MHz is ~500 ohms, so Zsec >> 500 ohms is needed), which is somewhat inconvenient at the high frequency here. On the upside, that's the magnetizing impedance; the characteristic impedance can be quite low, indeed should be, so that it is dominant capacitive, and acts in parallel with the electrodes.
On the primary side, you attach whatever matching impedance is needed. You will probably add additional capacitance in parallel to lower the frequency. A series matching inductor (thus making it an L match network) has a step-up feature which is valuable, though not so much with a transformer in use here. Alternately, a parallel inductor can be used to null out reactance, without providing matching. Note that a series inductor is still required for the class D amplifier (as its harmonics will not see a friendly impedance otherwise!), and may be desirable for the linear amplifier case.
Whatever the case, the tuning inductor should be adjustable as well, so as to keep Zo = sqrt(L/C) more or less constant, while varying L*C to set frequency. It won't actually be constant, because loss in the system will depend on frequency, but you can assist this by swamping system losses with an explicit resistor (which is probably a good idea, to set Q to a lower, more stable value).
Now, you might note the transformer has inductance, and that could be useful. But it's nonlinear (depends on core), and lossy (indeed, it may be the primary source of loss in the circuit). I would prefer using a transformer that's high enough impedance not to care (i.e., a transformer as such, as opposed to coupled inductors), and adding components around it to get the desired response.
You could just as well use an air cored (or well gapped, e.g., ferrite rod core) transformer, wound in much the same way but made larger to achieve the required inductance. It wouldn't be easy to adjust, though. Best way is probably an adjustable core, which might be ferrite to increase the inductance, or copper to reduce the inductance. I worry that this won't achieve enough adjustable range, or that the winding will be inconveniently large for the frequency range. (The adjustable range is about 2-4x for either method, or twice that if both are used alternately.)
Tim