circuit breaking: Change max_requests to work for both HTTP/1 and HTTP/2

[tonya11en]

The current way circuit breakers are used to limit outstanding requests involves using max_requests for HTTP/2 and max_connections for HTTP/1. This is unintuitive and and requires special handling for code that requires knowledge of the number of outstanding requests (such as retry budgets).

I'd like to propose altering the way we increment outstanding requests in the per-cluster resource managers to track requests between both HTTPs. This would allow users to simply set max_requests circuit breaker values and be agnostic to the HTTP version.

This is particularly risky because it's a subtle change in behavior for widely used parameters. Hiding this behavior behind a feature flag would certainly be necessary, but I'll defer to maintainers/community on the best approach to making a change like this (if it's desired).

[mattklein123]

+1 we should fix this. It has been a longstanding source of confusion. We definitely need to runtime guard this and probably initially default to off. cc @alyssawilk

[mattklein123]

Also related to #7403 in which we would optimally like to unify the H1 and H2 connection pool implementations as much as possible.

[oschaaf]

Drive-by comment:

This makes sense to me; a related idea after reading this:

The h2 pool implementation currently implies this, but once it supports multiple connections, perhaps a new max_requests_per_connection would make sense?
This could then translate into a (hypothetical) max request-pipelining depth for a semi-equivalent on h1.