The state ERROR can only be entered during decoder initialization, and
Seek() may only be called after initialization has been finished.

Fixes http://bugs.musicpd.org/view.php?id=4432

Allocate the buffer dynamically using av_malloc(), and free
AVIOContext.buffer in the destructor, as mandated by the libavformat
documentation.

Fixes http://bugs.musicpd.org/view.php?id=4446

Allocate the frame pointers only when actually decoding the file.

Eliminates some duplicate code, and as a side effect, this works
around clang 3.8 compiler warning because a negative value was
shifted.

This was lost in commit 8ead8f7e, because this was never implemented
for the new API.

Allow the compiler to invalidate buffer contents.

Move code from the "vorbis" decoder.

Improve initial seek by not reading/decoding the first frame before
checking for the seek command.

If there's a stream tag, don't let the song tag override it in the
next update_stream_tag() call.

If the song tag comes from a stream, and MPD playback restarts, MPD
would believe the tag should override the newly received tag.  This
makes the previous tag appear stuck.  This change passes the song tag
only if it's authoritative - i.e. if it's a song file.

Reduce runtime overhead.

When seeking to the beginning of a packet, skip the samples that come
before the desired time stamp.

Ask FFmpeg to seek to the next packet boundary *before* the seek
position, so we don't miss audio data.  Now we get too much, but we'll
solve that in the next commit.

We already know the offset.

Don't abort the process if there's not enough memory.  This buffer is
not important and can be large.

Merge two free() calls.