vf_overlay: rewrite request/push logic.

The old implementation, upon receiving a frame on the main
input, would request an overlay frame if necessary. This could
generate an infinite recursion, off-by-one frame mismatch and
other harmful effects, especially it the split filter is present
upgraph.

The new implementation uses the linear approach: it uses two
buffer queues for frames received out of turn and forwards
request_frame calls to the input where a frame is necessary.