Since ca068865c391e87932b1268d0c675be233dd2ffe the duration of the first
frame is not used for estimating the frame rate.

For this purpose, stream->first_duration was initialized with the
duration of the first frame. In fragmented files, this was previously
done by peeking the first moof, but that can only be done in pull mode.

Fortunately, we don't really need to do that, at least with the current
design: When we are estimating the frame rate we already have the
sample table, regardless of the scheduling mode and whether the file is
fragmented or not, so we can obtain first_duration there much more
reliably.

This fixes frame rate estimation for fragmented files in push mode.