AviSYnth can either return INT16, INT24, INT32 or FLOAT32 audio. I did some tests with it. When returning FLOAT32, the sound is almost the same quality as the input. Very little distortion at all. If instead using INT32, then the sound feels a bit more "square", losing a bit of smoothness.
As far as plugging into the ffdshow chain, is there any technical reason preventing from doing this in the same way as SVP is doing it?