os/bluestore: fix split vs finish_write race

os/bluestore: fix split vs finish_write race

In _tcx_finish(), we were looking at the right Cache for the collection,
and then calling finish_write with that Cache and taking the lock.  This
could race with a split_cache() such that after we got the lock the
collection was not on a different cache.  This would in turn lead to a
failed assertion later on in _rm_buffer when the sharedblob was trimmed.

Fixes: http://tracker.ceph.com/issues/24439
Signed-off-by: Sage Weil <sage@redhat.com>