If you use a bank of parallel transistors, I'd rather place MOSFETs instead of BJTs, because of the BJT's positive thermal coefficient (the more they heat, the more they conduct, so if one is more conductive, it will burn almost surely, followed by the others).
A good dummy load for measurement should be linear, not PWM, but it will be huge if power is quite high. PWM seems to me a good solution (especially if you plan to use a converter on the real system).
So, filter inductance * PWM period becomes the real matter. Higher values will make current more constant.
The issues with low inductance or low frequency values will be on the sensing side (current will be a heavy triangular wave added to a DC, but this can be filtered out), and on the measurement side, since the triangular ripple causes added losses (it has null mean value, but not null RMS value). The total load for the battery will be more than expected (you must take into account battery internal resistance and inductance ESR)...
This can or can not be a trouble, depending on the application. If you plan to convert battery voltage with a switching converter, a good thing could be to reproduce the same conditions in terms of ripple on the load, so the measure of duration will be very accurate.
Maybe I went a little off topic, or said thing you already know...