default flush timeout should be much lower, tunable to less than 1 second

[bcantrill]

When creating a Crucible upstairs, the flush_timeout is optionally specified in CrucibleOpts to be the number of seconds between flushes. Long flush times increase both space used by Crucible and time (as the number of outstanding jobs grows larger), but short flush times may result in excess CPU usage as fixed work is amortized over less time. When running measure-iops with the fixes described in #731 (implemented in #738), here is the effect of varying flush time:

The default of 5 seconds is too long -- it leaves quite a bit of performance on the table. Moreover, it is clear that there is a case to be made for a sub-second flush timeout. To avoid changing the units or semantics of flush_timeout, we propose changing it from an unsigned integer to a float -- and to changing the default from 5.0 to 0.5 (that is, 500ms).