hammer2 - Performance work
* Implement immediate (async) flush follow-through when OS requests
it, with write-clustering support at the device level if enabled.
Write-clustering support at the device level is disabled by default
because H2 already uses a fairly large and efficient 64K block size.
* Change the XOP dispatch for strategy functions. Further partition
strategy functions by read or write to prevent reads from getting
stuck behind long write chains.
Also do not span-out I/O related to a single inode across too many
XOP threads as doing so creates a massive amount of unnecessary
lock contention.
* Significantly improves concurrent filesystem ops and I/O and
significantly improves sequential filesystem write I/O.
UnifiedSplitRaw