Packet number recovery broken

[martinthomson]

I don't have a solid set of recommendations here, but after writing some tests, I think that the packet number recovery is broken. See this gist for some tests.

I can't really follow the logic that is there. It looks like it never rounds up, which causes it to pass the trickier underflow cases, but fail at a bunch of other tests.

I'm happy to provide an alternative implementation. The one from the spec is relatively easy to follow, or the shorter one. I thought that it might pay to discuss which is preferred before generating a PR.