The old flushing logic would simply leave frame threads (and tile
threads) running without caring how much latency that might impose
in the post-seek time-to-first-frame. This commit adds a 'flush'
state that will abort all running frame/tile threads from decoding
their current frame, as well as dispose of all frames in the output
queue.

Then, we use dav1d_flush() in dav1d_close() to abort running threads
on exit, instead of signaling their respective dependents to prevent
deadlocks. The advantage of this approach is that we don't signal on
objects we don't have ownership over, and thus this should prevent
race conditions where the owning thread could dispose of the object
just as we're signaling it, which I believe is what causes #193.