MC uses JRiver MP3 decoder, which is of very high quality (
http://yabb.jriver.com/interact/index.php?topic=37993.msg258591#msg258591). It has nothing to do with Winamp.
The MP3 decoder outputs 16-bit integral. MC can output all decoded formats bit perfectly, but the internal DSP works in 32-bit float. So if any DSP is used (including the Internal Volume and Replay Gain adjustments) the processing is done in 32-bit float. The selected output bit depth is applied after the processing.
I have done some testing with different decoders and I actually think that it may be better to decode lossy formats like MP3 to 16-bit integral. It looks like the peaks that decode over 0 dBfs with floating point decoders contain only encoding artifacts and these artifacts would actually be attenuated if not fully preserved. In any case this is a purely theoretical question because these peaks seem to always be too short for human hearing (below 1 ms).
CD-audio is 16-bit so it does not need more than that. In general, lossy formats do not preserve useful audio information that would need a greater bit depth (16-bit => 96 dB S/N ratio) even if encoded from a higher bit depth source. Everything that is assumed to be below the hearing threshold is already stripped away and the effective S/N is always much lower than 96 dB.