Proposal: SQLite multiprocess backend

[akx]

Following up on #367, I decided to try out whether it'd be feasible to add an SQLite backend for multiprocess values.

It's generally a little slower writing values, but a whole lot faster reading them out, especially when there are many PIDs! (Full disclosure, though: I'm only emulating multiple processes by hacking os.getpid() in my test bench, so these numbers could be off in a real-life situation.) See below for results.

⚠️ This hasn't yet been tested in a production situation at all, nor profiled – I'm sure there are some suboptimal bits and pieces in there. I felt like it would be better to post a work-in-progress PR anyway.
(Come to think of it, another avenue to investigate, if SQLite is considered too heavy-handed, would be some sort of append-only structure of files that can be easily collated and merged.)

[brian-brazil]

if SQLite is considered too heavy-handed, would be some sort of append-only structure of files that can be easily collated and merged.

That'd be even slower. I'd like to keep to no dependencies for this, as that makes it easier to adopt the library. Any tricks that sqllite is pulling, we can potentially pull too.

I'm suspicious of these results, I'd expect a lot more overhead from SQLite and the GIL is probably avoiding contention here.

[brian-brazil]

This has gone stale, and doesn't make sense to me.