feat: compute-aware ABR feedback for Shaka and dash.js plugins

[privaloops]

Adds a segment-level perf bus on which SegmentTranscoder publishes a
SegmentPerfStat (speedX = segDurMs / totalMs, width, height, frames)
for every successfully transcoded segment. A new ComputeAwareDecider
rolls those measurements through a hysteresis window and emits
subtractive cap decisions (lower / raise / hold). Player-specific
adapters wire the decider to each player's ABR knobs:

• Shaka: handle.attachComputeAware(player, opts?) -> player.configure
  with abr.restrictions.maxHeight + maxBandwidth caps
• dash.js: attachHevcSupport accepts adaptiveCompute -> updateSettings
  on a maxBitrate cap

ON BY DEFAULT. Bandwidth-based ABR has no way to observe transcode
cost, so a perfectly reachable variant can still saturate the device's
CPU. Default opt-in matches the actual cost shape of this library;
pass adaptiveCompute: false to opt out.

Defaults tuned for the compute-aware shape:

• measureWindow: 2 (was 8)
• lowerAfter: 1 (was 2) — buffer underrun is much worse than a brief
  unnecessary cap dip, so the decider reacts in ~2 segments instead of
  waiting for an 8-window average
• raiseAfter: 6, targetSpeedX: 1.3 (unchanged — still need real
  hysteresis on the raise side)

attachComputeAware(player, opts?) merges register-time options with
runtime options — convenient when the telemetry sink (onObservation)
is only known once the UI exists. onObservation now receives the
decision reason (init | hold | lower | raise) for richer diagnostics.

The decider snapshots state before each decision so applyCap failures
(player destroyed, etc.) can be reverted in-band. Cold-start guard
prevents raising before any lower has happened.

Both plugins re-export subscribeSegmentStat + SegmentPerfStat from
@hevcjs/core for custom telemetry on the raw perf bus.

Demos wired with a compute-aware live overlay (active quality + segment
speedX + cap + reason) and a diagnostic perf-bus probe. README +
per-package READMEs + site docs updated.

Closes #127

[github-actions]

Benchmark Results (GitHub Actions runner)

Resolution	FPS	ms/frame
1080p (50f)	36.5 fps	27.4 ms/frame
4K (25f)	10.4 fps	96.2 ms/frame

Note: GitHub Actions runners (2-core x86) are ~2-3x slower than Apple Silicon.
These numbers are for regression detection, not absolute performance.

_{Auto-generated by CI benchmark}