perf(pipeline): single-pass streamed peaks + presence

[thcp]

Summary

PR-3B of Phase 3 (Performance).

• New app/pipeline/audio_stats.py::scan_stem() does one streamed pass over a stem WAV via sf.blocks(), producing both the waveform peaks ([min, max] per bucket) and RMS (for stem presence) from the same blocks. Constant memory — a block is a few MB even for a 20-minute stereo stem, vs. sf.read()'s full in-memory load (~420 MB for the same file, done for up to 8 stems back-to-back right after Demucs has already stressed memory — a plausible contributor to OOM failures on memory-constrained machines like the reported 8 GB Mac Mini).
• collect.compute_stem_peaks now delegates to scan_stem and additionally returns each stem's RMS. peaks.json's format and bucketing are unchanged — floor-division chunking, matching the old implementation bucket-for-bucket (verified by a golden test comparing against the old sf.read()-then-chunk reference on a real multi-tone signal).
• runner._run_common now derives stem_presence from that RMS map instead of a second, separate decode. analyze.compute_stem_presence (which spawned its own ffmpeg downmix per stem) is deleted.

Closes #286
Closes #287

Known, accepted delta

Presence RMS is now measured over the full stem at full sample rate, vs. the old ffmpeg-downmixed mono decode capped at the first 180s. Sanity-checked against a real 220s track from a local job folder: this shifted some quiet-stem presence values by up to ~6 points (piano 1→6, other 12→19) — larger than the original estimate of "≲1-2 points," but a strict accuracy improvement (whole track vs. a 3-minute window), not a regression. Loud stems (bass, drums, vocals) moved by 2-6 points too, same direction.

Test plan

• New tests in tests/test_pipeline_collect.py: RMS matches a full-load np.sqrt(mean(x**2)) reference within tolerance; a stem missing from disk is excluded from the returned RMS map (same as peaks); golden test — streamed peaks match the old sf.read()-then-chunk implementation bucket-for-bucket on a multi-tone signal.
• New tests in tests/test_pipeline_runner.py for _presence_from_rms: normalizes to the loudest stem, empty input, all-silent input.
• Manually ran compute_stem_peaks against a real 8-stem job folder (220s track) — correct RMS values, peaks.json written, ~0.8s for all 8 stems.
• Full suite: 194 passed, 12 skipped
• ruff check / ruff format --check clean