Trace tooling

[mtrudel]

This is a WiP PR that adds some experimental tracing support to Bandit in order to facilitate easier diagnosis of ephemeral issues that users report such as #412, #431, #438, and really countless others before them. The idea is to provide a temporary 'trace' mode to Bandit which watches for exceptions (or user triggers) and when such a condition triggers, a dump of all relevant telemetry events is made. The hope is that this should help provide some insight into the causes of such issues, since the telemetry events contain a record of all bytes read/written, along with a bunch of other useful internal states (such as what the Conn & Bandit adapters looked like at each step of the way).

I'm planning on also adding some more untimed Bandit telemetry events at various places in the stacks (reading/writing of HTTP/2 frames, for example) to further enrich these dumps.

I'm not 100% sure that this approach will bear fruit, but there's no better way to try it out than in the real world. @atavistock, would you be able to take a look at the tooling on this branch and see if you'd be able to make use of it in the cases where #438 is happening? You can find useful docs at the top of the Bandit.Trace module in this PR.