There is no difference between pushing out a buffer directly
with gst_rtp_base_depayload_push() and returning it from the
process function. The base class will just call _depayload_push()
on the returned buffer as well.

So instead of marshalling buffers through three layers and back,
just push them from one place in handle_nal() and always return
NULL from the process vfunc. This simplifies the code a little.

Also rename _push_fragmentation_unit() to _finish_fragmentation_unit()
for clarity. Push sounds like it means being pushed out, whereas
it might just be pushed into an adapter.

This change has the side-effect that multiple NALs in a single STAP
(such as SPS/PPS) may no longer be pushed out as a single buffer if
we output NALs in byte-stream format (i.e. not aggregate AUs), but
that shouldn't really make any difference to anyone.