Problematic code path:
1. sps_buffers[0] is allocated in the first thread's codec context when the
decoder is opened.
2. The first thread context is memcpy'd to the other threads by frame_thread_init().
3. The first thread is closed and its sps_buffers[0] is freed.
4. The third thread is closed.
Because it never got to decode a frame, update_thread_context was never called,
and sps_buffers[0] still contained the first thread's pointer.

Fixed by not trying to free sps/pps buffers if the thread wasn't initialized.
I didn't properly consider this when designing it but this seems to be the
best approach anyway.

Fixes still2.mp4 crash from Chromium