We frequently had issues with notmuch segfaulting and my guess is that this
was due to the garbage collection magic used in the module.

This changes to a fork that ripped the functionality out.

There seems to be some race with the automatic closing that should happen
in theory... close it manually where we can to avoid the issue

Opening a notmuch DB gives you a snapshot of the stage at that specific time.
Prior to this, we only reopened the DB upon writing.
However, if say a mail sync program like offlineimap is fetching new mail,
we would never pick it up.

This commit caches a db for a while, so that we don't generate too much overhead
and does a reconnect cycle after that.

I hardcoded a value as I don't think that having an option would be beneficial.
Any write operation (meaning reading mail) anyhow flushes the DB by necessity.
(we need to close to commit tag changes, which changing the read state is)

Syncs back special notmuch tag like unread to the underlying maildir store

For some reason the current code frequently segfaults due to an
invalid C memory address. This commit mediates that by never keeping an object
alive longer than absolutely necessary.