Use separate mixing loops for the dry path and wet paths

Use separate mixing loops for the dry path and wet paths
This will increase CPU use a bit for sources that use auxiliary sends
(particularly with the cosine resampler), but it makes it more scalable when
adding more sends per source. Additionally, it will improve performance of
sources that don't use auxiliary sends.

The cosine resampler can probably be improved by using a lookup table instead
of calling cos() as there are 14 bits of fractional precision, so a 16384-
element float array would cover it all. This can also be reduced to a half or
quarter size if the precision loss is found to be acceptable.