It's supposed to be called only from (de)muxers.

The frame duration was calculated without taking in account the
bytes per sample.

Thanks to Lorenzo Pistone <blaffablaffa@gmail.com> for pointing
the issue and providing an initial fix.

It currently use the simple api and is using the latency information
provided only to offset the stream start.
Signed-off-by: Luca Barbato <lu_zero@gentoo.org>