This reverts commit 4f00910.

This ensures that the coordinator process flushes on demonitoring of the
attachment refs in chttpd_db. The problem here is that it's possible to
receive a 'DOWN' message for the monitor ref that is not receive'ed,
causing it to stick around in the coordinator message queue while the
next http request is handled. The pending message will not become
apparent until the next fabric call is invoked, as fabric expects to
have full access to all messages in the calling process, an expectation
which is violated by the pending message and causes a case clause crash
in the fabric receive message callbacks.

I noticed this during eunit runs with stubbed attachment handles that
generate an immediate noproc message on the monitor call. Normal
operations should not result in an immediate noproc result on monitoring
the attachment process, however, any failure that causes the attachment
process to fail between acquisition of the refs and the demonitor calls
will induce this bug, causing the next http request handled by the
particular chttpd coordinator pool processs to fail on whatever next
fabric call is made.

We need to potentially extract the usage delta from the incoming RPC
message. Rather than pattern match on all possible message formats that
could potentially include usage deltas, we instead utilize
rexi_util:extract_delta which matches against tuples ending in
`{delta, Delta}`, and separates that out from the underlying message.

The subtlety here is that receiving the message to extract the delta
changes the behavior as this was previously doing a selective receive
keyed off of the Ref, and then ignoring any other messages that arrived.
I don't know if the selective receive was intended, but I don't think
it's appropriate to leave unexpected messages floating around,
especially given things like issue #4909.

Instead of utilizing a selective receive, this switches to extracting
the message and delta like we need to do, and then in the event it finds
unexpected messages they're logged and skipped.

This selective receive was masking the lack of unlink on the linked
rexi_mon pid in fix #4906. I've also noticed some rpc responses arriving
late as well, but I haven't tracked that down, so let's log when it does
happen.

Most tests should not persist config changes. This is no exception.
While the test deletes the change afterwards, system halts between setup
and teardown result in a persisted max doc setting that causes
unexpected failures on the next test suite that writes a document larger
than 50 bytes.