feat(sound): add FineLAP audio event search

[tulayha]

Summary

• add a first-class FineLAP sound capability for music, environmental sounds, and general audio events
• store both global 10-second window embeddings and dense timestamped activations at 160 ms intervals
• expose sound search through the shared CLI, HTTP, MCP, browser, and Desktop contracts
• pin FineLAP and its RoBERTa text assets for reproducible local preparation
• refresh the multimodal model direction and benchmark documentation, with LongVALE retained as the next sound-quality evaluation target

Validation

• uv lock --check
• uv run --no-sync ruff check .
• uv run --no-sync python -m pytest -q — 706 passed, 5 skipped, 102 subtests passed
• npm --prefix desktop run check — typecheck, ESLint, 41 tests, and production build passed
• npx --yes markdownlint-cli2@0.23.2 "**/*.md" "#node_modules"
• real local FineLAP prepare, model-load, audio-indexing, and text-to-sound search smoke test

Notes

• FineLAP inference is local and does not require a paid model API.
• Sound retrieval quality has not yet been scored on LongVALE; the docs identify that as follow-up benchmark work.
• Link checking could not be run because Lychee was unavailable and the Docker daemon was not running.