Fix auto pong responses not flushing after block

[alexheretic]

Currently auto pongs are written and flushed during read(). If a WouldBlock error is received it is ignored and reading continues.

In the case the usage is read-predominant, this presents a scenario where the auto pong is WouldBlock-ed and will never retry flushing until another ping is received or the client does some unrelated writes. This could cause the other connected party to time out the connection as it has failed to receive a timely pong response.

I reproduced this in new test read_usage_auto_pong_flush.

To fix we record if a pong write failed in this way and retry on all subsequent read calls until a flush succeeds.

[daniel-abramov]

Good catch! Thanks!

NB I've now [again] realized that we need to rewrite our state machine (the code is getting increasingly convoluted and hard to manage) (edit: not the fault of this PR, just a general remark).