When sending original frames, do the reading in the queueing thread.
Earlier, we'd only preload it and hope the encoder's second read would
come from the cache. But this is more consistent, and perhaps more
importantly, it fixes a race on the FrameReader, where the two threads
would try accessing it at the same time, causing the file descriptor
to be closed while being used, and a crash.
Now, the FrameReaders are consistently accessed from one thread only.