feat(ltx2): Metal backend support (correctness first, then performance)

[jdilley]

LTX-2 hard-rejects Metal (Ltx2Backend::ensure_supported in crates/mold-inference/src/ltx2/backend.rs:34, as of 1503696): "not supported on Metal yet; use CUDA for real inference or set MOLD_DEVICE=cpu for correctness-only fallback". That means macOS users get no video from the flagship family, against the project goal of working on all hardware.

Suggested phasing

1. Correctness: enable Metal with the same status CPU has today (family capability contract row moves Metal from Unsupported to CorrectnessOnly), fixing whatever kernel gaps the candle fork has for the LTX-2 op set (attention chunking, FP8 → bf16 dequant path, audio VAE/vocoder ops).
2. Performance: adaptive offload tuning for unified memory, quantized matmul coverage, and a realistic default preset (e.g. distilled + reduced frames) so an M-series Mac produces usable clips.

Update docs/qualification/multi-gpu-family-matrix.md and the capability contract tests (crates/mold-inference/tests/family_capability_contract.rs) as the status changes.