eth,p2p: count timeout packet towards rtt

[holiman]

This PR changes the way we calculate RTT a bit. I noticed that for some nodes (particularly our azure nodes), there are a lot of dropped/failed trie heal messages.

Whenever we time out a packet, we update the capacity for that particular kind to 0. But what we do not do is update the rtt estimate. This PR changes that, so that a failed delivery does update the RTT, by pretending that it completed in 2x the amount of time spent so far. So if it times out after two seconds, it RTT estimate is updated as if it had been successfully delivered after 4 seconds. The idea is to make timeouts bump up the RTT estimate.

This is from one of the bootnodes, which is syncing. The graphs shows Unexpected trienode heal occurrences in the log. I restarted it with this PR at 11:50, when if flattens out. The dotted line represents 500 messages.

It's only one isolated and perhaps misrepresenting metric, but in that instance it seems to have improved the misdeliveries.

[holiman]

After some more time: this PR does not seem to make much of an impact on the log messages:

As for the grafana chart, I'm not sure how to interpret it:

[holiman]

Closing this. In case a couple of peers do not have the thing we're asking for, and legitimately never return responses, this would "indefinitely" keep degrading the rtt.