Implement 'notmuch dump'.

This is a fairly big milestone for notmuch. It's our first command
to do anything besides building the index, so it proves we can
actually read valid results out from the index.

It also puts in place almost all of the API and infrastructure we
will need to allow searching of the database.

Finally, with this change we are now using talloc inside of notmuch
which is truly a delight to use. And now that I figured out how
to use C++ objects with talloc allocation, (it requires grotty
parts of C++ such as "placement new" and "explicit destructors"),
we are valgrind-clean for "notmuch dump", (as in "no leaks are
possible").