Check for STOP before decoding the first chunk.  This reduces the
command latency.

This allows the compiler to discard buffer contents between two
iterations.

Nobody cares.

Prepare refactoring the whole plugin to use blocks instead of bytes.
A block is the smallest addressable unit, and it will simplify the
seeking code.

This finishes the multi-channel support.  Development of the feature
was started with commit 02cc77cd

When calculating the upper bound using the "sample count" format
header, don't assume it's stereo.

The plugin was horribly bugged for files that were not stereo.

Don't pass the buffer size to the function, as it's known at compile
time.  Use "restrict" on the pointer arguments, and merge the two
loops, which allows the compiler to optimize this loop with a few SSE2
instructions.

Read one block at a time.  This discards the last partial block, which
cannot be interleaved anyway.  Previously, uninitialised memory was
used to interleave the last block, which generated some noise.

Convert into a second buffer that gets passed to decoder_data()
without copying back to the first buffer.

This allows the compiler to discard buffer contents between two
function calls.

Reduce header dependencies.

If the offset is out of range, the seek will simply fail.  Not a
problem.

Merge lots of duplicate code.