Har-Bal 3 does all DSP using 64 bit floating point internally.
For real data streams there is essentially no material difference in performance between storing in 32 bit floating point or 32 bit fixed point.
On the issue of being able to disable the dither on 16 bit files, I don't see it as a good idea. For one thing, just because something has been dithered once does not mean it should not be dithered a second time if additional processing takes place. For example, if the dither in the original track is a noise shaped dither like super bit mapping and you make spectral modification in the normal audible range, not applying a dither will create significant harmonic distortion.
Basically, assuming that because a track has been dithered once will not need dither again with applied processing is flawed logic. For example, if you take the original source material and apply low level dynamic range expanding to reduce noise in silent parts, then the dither in the orignal track will not have sufficient level to function properly as a dither anymore, so you will have harmonic distortion introduce in those places where this processing is in effect. Similar things can happen with applied EQ as well so it is a flawed assumption to think that a dither is no longer required if one has already been applied. That is only true if you don't modify the track.
Having a choice between too much dither and not enough, the former is much better. The added noise of too much dither will be less audible than the distortion resulting from not enough so I won't have any plans to implement a mechanism to completely switch off the dither.