Play Framework

test(bench): add PF-134 SSE backpressure perf harness (micro + e2e, gated)

Opt-in performance harnesses; nothing here runs during `ant test`.
- framework/bench/sse: microbenchmark driving LazyChunkedInput's write/drain/
  concurrent hot path (standalone, built ad-hoc against framework/classes)
- SseLoadBenchTest: end-to-end HTTPS load test gated behind -Dsse.bench (forwarded
  to the integration fork via a conditional jvmarg); adds /benchSse + /heapUsed to
  the integration test app. Aborts via assumeTrue unless enabled.

Validates PF-134: memory bounded under a stalled client (~8 MiB vs unbounded
pre-fix), read path faster (zero-copy), and a ~16% single-stream firehose
throughput cost from the shared byte counter that only appears at unrealistic
event rates. Findings + sample numbers in framework/bench/README.md.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>