investigate: Possibly drain connections faster by not "re-draining" abandoned paths

[matheus23]

We drain paths on close that might've started "draining" a little longer ago for too long.
E.g:

• Establish connection at 0ms
• Open Path 1 at 100ms (it has the same RTT and variance as the initial RTT)
• Abandon Path 1 at 200ms (it now "starts draining" in the path-close sense, not the connection close sense. It will drain for 3 * PTO = 3s)
• Close the connection at 3000ms. Now we drain all paths. We also drain Path 1, which makes pto_max_path be 3s. So we drain for that time (even though we've started "draining" it 2.8s ago)
• Connection closes at 6000ms having drained fully.

Originally posted by @matheus23 in #366 (comment)

[flub]

So I played a bit with this. You need to get the last sent packet out of PacketNumberSpace::sent_packets, which shouldn't be that hard but needs some changes to the SortedIndexBuffer, and use it's SentPacket::time_sent for each path as base for the start of the 3 * PTO counter.

But like, still not sure this is logically the correct thing to do. Anyway, I've overspent my timebox on this for now.

[matheus23]

Tentatively put this under the 1.0 RC milestone so we check again if we want to do this before 1.0