We compute dump bandwidth basing on the time it takes a run writing task
to complete. While it used to work when we didn't have data compression
and indexes didn't share in-memory tuples, today the logic behind dump
bandwidth calculation is completely flawed:

 - Due to data compression, the amount of memory we dump may be much
   greater than the amount of data we write to disk, in which case dump
   bandwidth will be underestimated.

 - If a space has several indexes, dumping it may result in writing more
   data than is actually stored in-memory, because tuples of the same
   space are shared among its indexes in-memory, but stored separately
   when written to disk. In this case, dump bandwidth will be
   overestimated.

This results in quota watermark being set incorrectly and, as a result,
either stalling transactions or dumping memory non-stop.

Obviously, to resolve both issues, we need to account memory freed per
unit of time instead of data written to disk. So this patch makes
vy_scheduler_trigger_dump() remember the time when dump was started and
vy_scheduler_complete_dump() update dump bandwidth basing on the amount
of memory dumped and the time dump took.