The previous implementation assumed that a new picture would always
supersede the previous picture. Similarly, presentation segments
were assumed to pertain to the most-recently-read picture.

However, each presentation segment may refer to 0 or more pictures
by their ID. Picture IDs may repeat, and a repeated picture ID
indicates that the old picture for that ID is no longer needed
and may be discarded.

The new implementation allocates a buffer with one slot for each
possible picture ID (the picture ID is a 16-bit field) and
properly decodes presentation segments so that all relevant
pictures are output upon encountering a display segment.

Given that most PGS streams are unlikely to use more than a small
fraction of the available picture IDs, it would probably be better
to use a more memory-efficient data structure. I'm lazy though, so
I leave this to a more motivated individual.

I've tested the code with MKV files in VLC (a recent revision from
their git repo) and with HandBrake (a version that I hacked up to
use ffmpeg's PGS subtitle decoder).
Review-by: Hendrik Leppkes <h.leppkes@gmail.com>
Signed-off-by: Michael Niedermayer <michaelni@gmx.at>