perf(reactor): remove the late-round heuristic (CWIST_REACTOR_LATE_US)

[gg582]

Summary

Removes the late-round heuristic (CWIST_REACTOR_LATE_US, default 3000 us) that bc358b4 added as a #319 experiment.

The heuristic watched each io_uring dispatch round. Once a round had run past the threshold, it switched the rest of that round into late mode:

• each callback's re-arm was flushed to the kernel right after the callback;
• posted work was drained after every callback;
• pipelined HTTP/1.1 connections served their buffered requests without yielding.

Why remove it: the same-runner A/B never showed a tail improvement beyond noise. The follow-up age-based version (#375, now closed) was neutral up to c1024.

Removed

• src/sys/io/reactor.c:
  • reactor_late_ns() and cwist_reactor_dispatch_late();
  • the per-callback clock read and the late flush/drain;
  • the late_rounds / late_callbacks latency-probe counters.
• src/sys/io/reactor_rx.h: the cwist_reactor_dispatch_late() declaration and its non-Linux stub.
• src/sys/app/app.c:
  • the HTTP handler goes back to yielding every yield_batch requests, with no late bypass;
  • the now-unused reactor_rx.h include is gone.
• tests/test_reactor_late.c and its Makefile target.
• The CWIST_REACTOR_LATE_US rows in docs/runtime-config.md and docs/ko/runtime-config.md.
• The "CWIST C1M (late mode off)" benchmark leg in bsd-kqueue-benchmarks.yml, and its mention in scripts/ci/benchmark.py.

Unchanged

• CWIST_REACTOR_ROUND_BUDGET_US (round-budget flush) and CWIST_REACTOR_DRAIN_CHUNK (post-drain chunking).
• Past cwist_c1m_lateoff_* rows in benchmarks/webserver.json, kept as history.
• Public API: this touches only internal headers. check_api_baseline passes.

Verification

• Full local make test passes on both the default backend and CWIST_REACTOR_BACKEND=epoll.
• check_test_wiring, check_env_knobs, check_api_baseline and check_inline_exports pass.
• The workflow YAML parses.