number of channels is specified when the converter state is created.
Previously this was only done once, thus breaking horribly when the input
audio suddenly had a different channel count.  A new state could be created
every time the number of channels changes, but this could happen many times
a second if resampling to two different formats at once.  The simple API
doesn't have this problem, as channel count is an argument to the
conversion function itself.

git-svn-id: https://svn.musicpd.org/mpd/trunk@6229 09075e82-0dd4-0310-85a5-a0d7c8717e4f