test: harden DrumSep e2e clip + reconcile MDXC chunk-starts (#298×#299) (#303)

* test: harden DrumSep e2e with a content-rich drum clip

The DrumSep test used levee_drums.flac, whose drums contain no ride-cymbal
content — so the DrumSep ride stem came out silent and its correlation-vs-
reference check compared two near-silent tracks, yielding meaningless ~0.47
noise that flip-flopped around the 0.70 threshold (this false-positive briefly
blocked #299).

- Add tests/inputs/drumkit_groove.flac (25s isolated drum-kit performance) so
  every kit part has audible content.
- Run DrumSep directly on it (the clip is already drums; htdemucs drum
  extraction is covered by test_four_stem_separation) and regenerate refs.
- Add a silence-floor tolerance: stems whose reference is near-silent only
  require the output to also be near-silent, since correlating silence is
  not meaningful.
- Update generate_multi_stem_references.py to match; drop orphaned
  ref_levee_drums_*_drumsep.flac.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

* test: reconcile MDXC roformer chunk-starts assertion (#298 × #299)

test_step_uses_overlap_divisor asserted chunk starts [0, 4, 8, 12], the
pre-#298 schedule. #298 rewrote _roformer_chunk_starts to stop once a chunk
reaches the input end, dropping the redundant offset-12 tail chunk (which
clamps back to start 8). Each PR was green on its own base, but the combined
main was never CI-tested together (the release version bumps went straight to
main), so this semantic conflict only surfaced now.

The runtime behavior is correct and better (no wasted tail chunk); update the
assertion to [0, 4, 8], which still demonstrates the step=chunk_size//overlap
divisor semantics via the step-4 spacing.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

---------

Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>