Network worker might miss wake ups

[tomaka]

See the discussion here: #6247 (comment)

This is probably not very important because under heavy load the network worker will wake up anyway at some point, but should still be investigated.

[romanb]

I looked into this a bit and from what I could tell, the large variations in runtime of the test from that discussion is not due to missing task wake-ups but rather because connections often get closed due to an upgrade failure in the ping protocol when reading the response. Ultimately the problem seems to be libp2p/rust-yamux#81 and the fact that the outbound upgrade of the ping protocol issues an early half-close of the substream exacerbates this problem, i.e. makes it much more likely to occur. In any case, with the linked patch the test takes consistently ~6 seconds for me with 32 peers and ~9 seconds with 256 peers (considering that the test has an artificial 5 second delay for any execution, that seems ok), without seeing any unexpected errors or closed connections in the logs.

[twittner]

I think the problem with ping is that this socket is not closed.

[romanb]

Feel free to reopen if you think this still isn't resolved.