Add X-LANCE multi-stem splitting

[Gabriel-Kahen]

Summary

• add an X-LANCE Stem Splitter action to audio-asset context menus and the asset inspector
• expose all eight X-LANCE targets plus denoise and vocal-dereverb quality options
• show per-stem progress and support cancellation of the complete inference process tree
• import completed stems as one undoable group under SPLIT - <original file name>, with children named <INSTRUMENT> - <original file name>
• persist asset folders in the canonical project format and render them as virtualized, collapsible groups in the asset sidebar
• support the bundled adapter on Linux with an NVIDIA CUDA GPU using a pinned, app-managed X-LANCE checkout and Linux dependency lock

Review fixes and quality safeguards

• make project completion delivery lossless and prevent queued GUI edits from overwriting an imported stem group
• reject project races, changed source bytes, stale revisions, wrong duration/sample rate/layout, non-finite output, and incomplete output before canonical registration
• split the imported asset's actual base media even when it also has a derived current revision
• keep unrelated empty asset folders intact across deletion and undo/redo
• use project-local temporary storage, perform a free-space preflight, and reap the entire subprocess tree on cancellation or timeout
• serialize and clean the app-managed X-LANCE checkout so concurrent jobs cannot race or execute leftover untracked code
• reduce long-recording memory use with float32 decoding and incremental overlap-add while preserving song boundaries
• use explicit high-quality SoXR resampling and restore exact source length and mono/stereo layout; mono results use a real stereo downmix
• version the asset index independently so older builds reject folder-bearing documents instead of silently discarding organization

Testing

• cargo test --workspace
• cargo clippy --workspace --all-targets -- -D warnings
• cargo fmt --all -- --check
• uv run --isolated --no-project --with numpy==2.2.6 --with soundfile==0.13.1 python -m unittest scripts/test_gaw_xlance.py
• adapter CLI syntax/help check and git diff --check

A real checkpoint inference run was not possible in this environment because it has no NVIDIA CUDA device. The subprocess contract, cancellation, project/storage flow, race handling, overlap-add behavior, and output normalization are covered without the model weights.