This is a respin of Dave's xfs_sync_inodes refactoring plus a couple of new
patches. The major difference to the original series is sorting out the
pag_ici_lock locking which was busted in many ways. We now held the lock
exactly as long as the current code, to do that I had to give up on the
lookup vs execute split for the callbacks which actually improved the
code flow a bit, at the expense of exposing the lock to the callbacks.
The other change is that inode flushing still uses the same old algorithm
instead of bringing over the optimizations from xfs_fs_write_inode which
caused some spurious regressions in test 183 for me. I still plan to get
these two paths unified, but let's leave that for a separate patch.