fix(http): classic shutdown waits for every handler before the app goes

[gg582]

Follow-up to #351 (the open TSan report in test_http_client).

Bug

Classic pool handlers run on detached workers. A keep-alive connection keeps its worker in a blocking read for as long as the client holds it open (the BDR fast path's recv has no timeout at all), and a client that keeps sending requests keeps it dispatching. Shutdown waited for the in-flight count only up to the drain timeout, then cwist_app_destroy() freed the route table under a worker still in cwist_route_table_lookup.

Fix

Each pool worker registers the connection it serves; cwist_http_pool_destroy() drains them before it returns (and before the caller destroys the app):

1. idle keep-alive connections are shut down at once;
2. in-flight requests get the drain timeout (CWIST_DRAIN_TIMEOUT) to finish;
3. what is left is shut down (SHUT_RDWR) and the pool waits until every handler has returned. A handler stuck in route code is waited for (warning every 5 s) rather than freed under, the same contract C1M already has (it joins its handler threads).

• Handlers mark the wait for the next request with cwist_http_live_idle_enter() / _leave() (private src/net/http/http_live.h): the header wait in cwist_http_receive_request() and the BDR recv in cwist_app_http_handler(). Once the pool is stopping, idle_enter returns false and the handler closes instead of waiting. Outside a classic pool worker both are no-ops.
• The handler serves a dup of the socket; the pool keeps the original open until the registry entry is unlinked, so a shutdown() issued under the registry lock never reaches a reused descriptor. Cost: one dup + close per classic connection (not per request), and one extra descriptor while it is served.

Scope: the plain-HTTP classic pool. The HTTPS pool and the forking path are unchanged.

Tests

• New tests/test_classic_shutdown.c (classic, one worker, 4 s drain):
  • idle keep-alive connection: shutdown in 0.02 s (before: 4.01 s, the full drain, then the app was destroyed under the worker — verified by running the test against dev);
  • in-flight /slow (1 s handler) still gets its complete response;
  • a client hammering its keep-alive connection cannot keep the server alive; every response it got was complete.
• test_http_client removed from TSAN_EXCLUDE.
• Under TSan, 3 runs each: test_http_client, test_classic_shutdown, test_listen_ex, test_classic_pool_scaling, test_shutdown — all pass. Normal build: the same plus test_static_and_range, test_async_defer, test_https_metrics pass.
• malloc_lint.sh, check_api_baseline.py, check_test_wiring.py OK (no public API change).