Return only -1,0,1 and not directly subtraction, as like Remi pointed
out, result can exceed INT_MAX.

Xml reader reads stream so we need to handle that so later on probing
don't fubar. This seemed to hit for srt probing atleast as currently it
doesn't anymore explicitly seek to 0.

Most likely no major speed difference with low amount of subtitles.

This allows deployment in multi-threaded OpenGL scenarios without collisions

regression since using absolute numbering

Tizen audio_io.h supports only mono/stereo, with PCM U8/S16.  With Tizen 2.3,
flush is done by destroying/recreating the audio_output. A real flush will come
with Tizen 2.4.

There is no TimeGet for now since there is no way to know the position in audio
stream.

The minimum SDK version supported is 2.3.

Also fixes lot of spurious decoder errors when stopping due to
excessive flush checks.

This variant of picture_pool_Get() sleeps until a picture is available.

This reduces the semaphore contention scope.

This allocates a picture without affecting the video output format.
This can be used when the decoder keeps proper track of format updates.

The coded video size is already fixed by the decoder when creating the
vidoe acceleration plugin, while the chroma is fixed by
vlc_va_GetChroma(). This callback is useless.

This call was originally meant to adjust the hardware acceleration to
video size changes, but that cannot happen there anymore. Also errors
cannot be handled properly there either.

This should never happen, and it would not work. We need the real
dimensions to initialize the video output thread.