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.

Same bug as in the previous commit.

When the data chunk size is not a multiple of the frame size, the last
partial frame lead to an endless loop.  We fix this by checking
chunk_sze>=frame instead of chunk_sze>0.  This way, the partial frame
is simply skipped.

Reduce header dependencies.

Seeking on NFS or SMB is cheap.  Actually, only HTTP streams are
expensive to seek.  This enables a few features on NFS/SMB files, for
example Ogg tags.

Has been moved to class InputStream long ago.

Since we already checked InputStream::KnownSize(), we can assume that
GetSize() returns a valid value, and this check is obsolete.

Replace the bogus GetSize() check and call GetSize() only when
necessary.

Variable-length arrays are not allowed in C++.

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

The plugin assumes that the size is known, but does not verify it at
runtime.

Generate more detailed error messages.

Eliminate duplicate definition (in input plugin and decoder plugin).