Add macOS Apple Silicon/MPS support with CUDA-parity export

[jgranie]

Summary

This PR adds a macOS Apple Silicon execution path for Pixal3D, validated on an M3 Max with 36 GB unified memory. It keeps the original CUDA path available while adding MPS-compatible attention/convolution fallbacks and a memory-bounded CUDA-parity exporter.

What changed

• Add macOS setup, dependency, diagnostics, and smoke-test scripts.
• Route device selection through MPS with CPU fallback where required.
• Add MPS implementations for sparse attention, sparse convolution, NAF neighborhood attention, and grid sampling.
• Add a chunked Metal BVH path for the 18M-face source mesh to avoid the monolithic BVH coverage failure.
• Add bounded remeshing, cleanup, texture projection, alpha validation, checkpoint resume, and GLB quality comparison tools.
• Document the validated workflow in README_MACOS.md.

Validation

• Test image: output/inputs/0_img_2048.png, seed 42.
• MPS final: 989,941 faces, 4K texture, opaque single-sided material.
• CUDA reference: 937,343 faces, 4K texture, opaque single-sided material.
• Dominant connected surface: 99.46% MPS vs 99.47% CUDA.
• Blender comparison completed from front, back, side, and underside views with no visible point-cloud, transparency, or internal-hole artifact.
• Python syntax checks and whitespace validation pass.

Performance note

The validated MPS generation takes about 35 minutes end-to-end on the M3 Max 36 GB, compared with about 13 minutes 25 seconds on an RTX A5000.

The RunPod CUDA archives are intentionally not included in this PR.