Parsing PAT resets PMT buffer. If PMT does not fit to single TS packet and
PAT packet is scheduled in middle of PMT packets, PMT is never parsed and TS
demuxer falls to PID auto detection mode.

This moves PMT buffer reset to the case where PMT PID changes and all PIDs
are reset. [As far as I can see, worst regression can be invalid PMT
sections when stream changes, new PMT pid equals to old one and demuxer is
not reset(?), but this should be OK as PMT checksums are always checked.]

There are two tricks to make VC1 decoding work:
1) VC1 sequence and entry point headers must be present in context->extradata.
2) video width and height must be known when opening decoder.

Some container formats store required extra data, but mpeg-ts does not.

1) is fixed by scanning the stream for headers and discarding all data until
proper headers are found.

2) is fixed by re-opening decoder with width and height information from
first open.