Peaks computation loads entire WAVs into RAM #286

Location: collect.py:198
Severity: HIGH

sf.read() materializes the full file: a 20-minute stereo stem is ~420 MB of float32, done for up to 8 files back-to-back — right after Demucs has already stressed memory. On an 8 GB Mac Mini this is a plausible contributor to the failure reported.
Stream with sf.blocks() (fixed 1500-bucket min/max fold) — constant ~few-MB memory, same output.

From the health report. Part of #274 (Phase 3 — Performance).