Just think about the transformer and winding resistances:
Apply 120V to the "near" primary (denoting the primaries as the 120V windings). You get 48V on the near (same-side) secondary, and slightly less than 120V and 48V on the opposite windings (the loss is due to leakage flux, which acts to slightly reduce the voltage ratio, among other things).
Apply a load to the near 48V. Voltage drops across its resistance, and the primary's resistance. (The voltage on the opposite windings drops slightly further, due to the primary IR losses only.)
Apply a load to the far 120V winding instead. Voltage drops across its resistance, and the primary's resistance, and the leakage flux (which manifests as a series inductance). The near 48V winding is slightly lower in voltage (due to primary IR loss only), and the far 48V winding is lower still (due to leakage and primary IR loss).
Leakage between windings on opposite sides of a core / magnetic loop will be fairly high, enough to be noticeable (affecting regulation) even at line frequency. But the windings within each core leg will be pretty well coupled. So you can treat each side as a sub-transformer of its own. By wiring the 48V secondaries in parallel, you are shorting out that leakage path between those sub-transformers. It's like using a pair of 120:48V transformers back-to-back to get 120 again. Except it's better, because the primaries are already mostly coupled (>90%, I would guess), so the 48V coupling only has to mop up the excess (the leakage, <10%).
With the 48V windings in parallel, apply a load to the far 120V winding. The voltage doesn't drop as far, because most of the leakage is shorted out. Now only the IR losses of each 120V winding remain. A small current flows between the 48V windings, due to flux balancing; I doubt this will be more than 10% of their rating, even when the far 120V winding is heavily loaded.
How much IR loss are we talking?
When the transformer is operating as designed, both 48V windings are powered, and the 120V windings are loaded (in parallel, I assume, unless they were used for 240VCT; in which case, I'll assume parallel equivalent anyway, for convenience). Which works the same as what we're doing, powering the 120V and (maybe) loading the 48V.
With all four windings active, each primary gets one unit of I*R, and each secondary gets one unit of I*R. The pairs of primary and secondary windings are wired in parallel, halving each pair's total resistance. The primary and secondary act in series, bringing the total back to one unit of I*R.
Likely, it was designed so that I*R is 5 or 10% of nominal output, i.e., for a 120V output (in the step-up configuration), it drops 6-12V at full load current (3kVA --> 25A). That's 12.5A per 120V winding.
And total winding loss was I^2*R. To leave the 48V windings unused, you drop half the loss, which is helpful, but you can't double the 120V windings' currents to compensate: that would quadruple their losses, and the transformer would run way too hot. At best, sqrt(2) more current can be drawn, but this has two problems: one, the winding internal (hotspot) temperature will be ~41% higher; and two, more current simply drops more I*R voltage, worsening regulation.
It's also slightly worse due to the flux balancing current. If the load is 120V and 12.5A, and the flux imbalance is ~10% like I worst-casetimated earlier, then the 48V windings will be carrying about 10% of their rating, i.e., 3A (out of 31A each). Which dissipates a little heat, so the best case will actually be like 10% below sqrt(2). But still, probably limited by regulation (how low can your line voltage drop, before it's really too bothersome to use on heavy loads?).
In any case, that's still 1.5kVA or so worth of capacity, which is not bad at all.
Tim