Waiting for the next NAL increases the latency. If alignment=nal/au
has been negotiated, assumes the the buffer contains a complete
NAL and don't except a second start-code. This way, nal -> nal,
au -> au and au -> nal no longer introduce latency.

As a side effect, the collect_pad() function was not able to poke at the
following NAL. This call is now moved before processing the NAL, so
it's looking at the current NAL before it's ingested into the parser
state in order to dermin if the end of an AU has been reached. The AUD
injection state as been adapted to support this.