perf: isolate macOS render and decoder pressure

[tcconnally]

Summary

• Coalesce libmpv frame-update notifications before they enter the Qt GUI queue.
• Export bounded frame-pump counters separately from render counters.
• Export requested/active decoder and per-resource fallback/quarantine state.
• Add a sanitized powermetrics/sample parser and bounded profiling CLI.
• Add the macOS M5 performance roadmap and link it from the README.

Why

The recent M5 evidence showed a real CPU/render problem, but not a decoder-only fix. The Hyperwall process averaged 2.14 CPU cores in the captured hardware trial. The corrected normalized render CPU was about 0.8275 core-equivalents for software versus 0.8313 for hardware. The frame pump therefore needs direct measurement and bounded queue behavior before another long soak.

Validation

• python3 tests/run_all.py passes with all repository suites green.
• Focused frame-pump, telemetry, decoder-observability, roadmap, and profile tests pass.
• The profile CLI was exercised against the real 600-sample powermetrics artifact and a sanitized stack capture.
• git diff --check, Python compilation, and shell checks pass.

Scope boundary

This PR does not claim that 8-cell playback is production-ready. It does not make videotoolbox-copy a global default and does not use the rejected runtime 30/25 live-transcode shaping. A short native M5 profile and measured 4/6/8-cell capacity gate remain required.

Tracking

Refs #75
Refs #76
Refs #77
Refs #78
Refs #79