mod_proxy: Add a per-worker maxbusy limit on concurrent requests, glo…

[arturobernalg]

max= on a BalancerMember only bounds the per-child connection pool; nothing today provides a process-global limit on concurrently active requests to a backend, so under multiple children/threads a backend can receive far more concurrent requests than an admin configuring max=N would expect.

This adds maxbusy=N to BalancerMember: a hard cap on requests concurrently active against that specific member, counted across all child processes and surviving a graceful restart. 0 (the default) keeps current behavior.

Design: a new admission table is kept separate from worker->s->busy (which stays exactly as it is today, including its use by bybusyness and its reset() semantics) and separate from proxy_worker_shared, which is per-restart-generation shared memory and so cannot hold live cross-generation state. The table lives in an apr_shm retained via ap_retained_data_get()/_create() (already used in this file for proxy_start_time), owned by the parent process, so it survives both a graceful restart and the unload/reload of mod_proxy.so that a graceful restart causes. Every statically configured maxbusy member is known by post_config time (the balancer-manager dynamic-worker form does not expose maxbusy), so every admission slot is pre-registered once in mod_proxy_balancer's post_config, after balancer->s->sname is final; the request path only ever looks up an existing slot and does an atomic CAS reserve/release, never creating one, so there is no creation race between processes or generations. A saturated member is treated like a disabled/error member for selection purposes (sticky redirect, hot spares, balancer timeout), not marked in error itself. Reservation release is a separate cleanup from the busy-count cleanup, run immediately when proxy_handler() abandons a worker for a failover attempt, rather than only at the end of the whole request.

Tested manually with real multi-child builds (prefork/worker/event), real graceful restarts with in-flight requests, BalancerPersist on and off, sticky/redirect/hot-spare/standby topologies, failover chains, and backend HTTP/2 and WebSocket/tunnel traffic. Added test/modules/proxy/test_07_maxbusy.py.

MMN bumped to 20211221.34 for the new proxy_worker_shared::maxbusy field and the new ap_proxy_* exports.

This PR does not report or address a security vulnerability, follows the httpd style guide, uses an empty APLOGNO() tag for its new log messages, includes a changes-entries/*.txt file, and includes pyhttpd test cases in test/modules/proxy/.