perf(hdr): reduce layered composite overhead

[vanceingalls]

Summary

• instrument HDR layered compositing with per-stage timing buckets
• switch HDR video frame extraction from PNG sequences to raw rgb48le files so the hot path does fixed-offset reads instead of PNG decode
• skip exact-zero-opacity HDR blits and use a faster byte-level opacity blend path for rectangular rgb48le layers
• add alpha blit coverage for opacity blending and zero-opacity no-op behavior

Performance

Local HDR benchmark (bun run benchmark -- --tags hdr --runs 1):

• before: 368.7s combined (hdr-hlg-regression 102.6s, hdr-regression 266.1s)
• after: 281.8s combined (hdr-hlg-regression 67.9s, hdr-regression 213.9s)
• net: ~24% faster combined

Validation

• bun run --filter @hyperframes/engine typecheck
• bun run --filter @hyperframes/producer typecheck
• vitest run src/utils/alphaBlit.test.ts
• vitest run src/services/renderOrchestrator.test.ts
• bunx oxlint packages/engine/src/utils/alphaBlit.ts packages/engine/src/utils/alphaBlit.test.ts packages/producer/src/services/renderOrchestrator.ts
• bunx oxfmt --check packages/engine/src/utils/alphaBlit.ts packages/engine/src/utils/alphaBlit.test.ts packages/producer/src/services/renderOrchestrator.ts
• git diff --check

[miguel-heygen]

Approved after reviewing the HDR compositing changes on head 9ce0da2. I checked the raw rgb48le extraction/read path, opacity blit changes, cleanup behavior, and perf reporting, and validated with targeted unit/type/lint checks plus the HLG HDR regression fixture. CI is green, including the HDR regression shard.