Progress estimation is usually based on transfer speed. That makes no
sense when we're doing operations like deletes, that need very little
data transfer but nevertheless take a long time.
This hack attempts to detect this case better and switches to a
different estimate.
We should rewrite this to maintain and update estimates for the
transfer speed, per-file overhead and chunk-assembly overhead each
time an item finishes. Then we could provide more consistent progress
estimates without ad-hoc fixes like this one.
Also, there's an issue where resuming a partial download will lead
to exaggerated transfer speed estimates.
