watch.  This should fix the race condition that seems to have caused
Issue #126: When another thread is buffering an insert for the record
that is being purged, it should invoke buf_pool_watch_notify(), so
that the purge will not be buffered. Otherwise, the purge would be
buffered for the wrong record (one that was inserted after the purge
determined that the record can be removed).

When deletes are not buffered, the latch on the secondary index page
would prevent the insert from occurring.  In delete buffering, the
buf_pool_watch replaces the page latch when the page is not in the
buffer pool.

buf_pool_watch_notify(): Make public.

ibuf_insert(): Invoke buf_pool_watch_notify() when buffering an
insert or a delete-mark operation.