Reuse idle timeout stage

[RafalSumislawski]

This PR:

• addresses the Possible Blaze Client Performance Regression in 0.21.23 #4916, in which a leak of idle timeouts / IdleTimeoutStage's was observed. Now there is zero or one IdleTimeoutStage per Http1Connection. That stage is created together with the connection and keeps being reused though the life of the connection. The logic of IdleTimeoutStage makes sure that at most one timeout is scheduled.
• avoids the cost or creating IdleTimoutStage per request. In many cases that cost is minuscule. But that's not the case when using log4j2 on Java 11 (see similar issue in armeria Performance degradation when using JDK 11 and log4j2 line/armeria#2306).
• Removes ReadBufferStage which was supposed to address the same issue as Read from an idle blaze-client connection to detect when server closes it #4810, but apparently it didn't work. And since Read from an idle blaze-client connection to detect when server closes it #4810 is merged the ReadBufferStage it obsolete.

[djspiewak]

Can confirm that this appears to fix the observed performance regression, at least to within a moderate confidence interval. There's a decent chance that performance is still slightly worse across the board around 15% slower in all percentiles, but more testing would be required to say for sure. At the very least, this is a really nice fix!

[Baccata]

We've tested it and are happy with the results 👍