The existing logic was not working sometimes because:
1. The repeated scheduling in releaseAllFramesFromMediaCodec was
starving the thread on which the SurfaceTexture frameAvailableListener
was called.
2. The case where a pending frame arrives on the surface after flush
finishes executing was not handled.
The consequence of both problems is that availableFrameCount ended up
being > pendingFrames.size().
PiperOrigin-RevId: 668916256
37561c829f