perf(http): single Map probe in read_headers via update_or_default

[mizchi]

Summary

Reader::read_headers in src/http/parser.mbt does two Map operations per header:

match headers.get(key) {
  None => headers[key] = value
  Some(value0) => headers[key] = "\{value0},\{value}"
}

Map::get walks the Robin-Hood probe sequence once; the subsequent headers[key] = ... walks it again from scratch. Map::update_or_default(key, default, update_fn) already exists for exactly this pattern and does a single probe.

headers.update_or_default(key, value, value0 => "\{value0},\{value}")

Why bother

Headers are parsed for every HTTP request. The save is small per call (one extra hash + one probe walk) but compounds across the request lifetime — and the new version is semantically identical and arguably clearer.

Scenario: bench-async/cmd/http_server_benchmark/main.mbt — the example HTTP server driven by k6 (3,330 req/s under valgrind, 67 kreq/s native). In the callgrind profile read_headers itself is 2.40% of instructions and the Map insert path inside it is a slice of that. Wall-time delta is within the run-to-run noise (~±2%) on a GET / workload — the visible work is dominated by moonbit_drop_object + malloc — but the patch costs nothing.

Test results

moonbitlang/async/http                  51 / 51 pass
moonbitlang/async                       87 / 87 pass
moonbitlang/async/aqueue                35 / 35 pass
moonbitlang/async/cond_var               5 /  5 pass
moonbitlang/async/semaphore              6 /  6 pass
moonbitlang/async/internal/coroutine     1 /  1 pass
moonbitlang/async/internal/event_loop    7 /  7 pass

Network round-trip tests (real socket / tls / http) require a network interface this sandbox lacks; this patch does not touch any code on that path.

[Guest0x0]

lgtm except the minor comment issue

[Guest0x0]

the comment feels unnecessary here