Add PyTorch MRT2 validation harness

[brxs]

Part of #109.

What changed

• adds an isolated two-deck PyTorch MRT2 validation harness without changing production code
• compares one shared worker with two independent deck states against one worker per deck
• exercises LSDJ's 25-frame/~1-second and 5-frame/~200-ms scheduling modes
• simulates the existing 1.5-second prebuffer and labels simulated underruns separately from engine telemetry
• records p50/p95/p99 latency, startup/shutdown, RSS, CUDA allocation, and NVIDIA telemetry
• pins the public upstream source and model/processor snapshots immutably and operates offline after acquisition
• documents the API/parity/packaging findings and a Linux/Windows NVIDIA qualification checklist

Current decision

This is a conditional no-go for production issue #110 until:

1. the 10-minute two-deck matrix passes on representative Linux and Windows NVIDIA hardware;
2. Add third-party model licensing, attribution, and download acknowledgement #108 resolves attribution/redistribution metadata for the re-keyed PyTorch weights versus the underlying CC-BY-4.0 weights and MusicCoCa assets; and
3. the packaged runtime proves its locked acceleration path does not need a compiler or mutable network fetches at runtime.

The canonical public upstream repository is available and supports a thin pinned adapter; the retired authentication-gated Space is not required.

Validation

• Ruff format and lint checks
• six unit/dry-run tests
• provenance JSON validation
• full synthetic matrix: two topologies × two frame modes
• git diff --check

Synthetic output is explicitly marked as non-qualification evidence. No hardware result is claimed by this PR.

[brxs]

Superseded by and fully consolidated into #131. The combined branch includes this lane's substantive changes, subsequent integration fixes, and exact-head cross-platform validation. Please continue review in #131.