patches/mps_compat: re-enable decode-time cumesh on Apple Silicon (closes the see-through-mesh bug)

[NoahBPeterson]

Summary

Removes the unconditional skip of cumesh's decode-time fill_holes /
remove_faces / simplify on Apple Silicon. The skip was a workaround
for a real bug — Metal cumesh used atomic_min / atomic_max on float,
which is an Apple9-only feature that crashed on M1/M2 — but the
workaround has been measurably hurting output quality.

With pedronaugusto/mtlmesh#1 in place (Apple7/8 atomic-fallback for the
simplify and atlas kernels), cumesh runs end-to-end on every Apple
Silicon GPU family. There's no longer a reason to skip these ops.

Why this matters

Skipping fill_holes specifically had a visible cost. On the reference
T.png input on an M1 Pro:

• Decoder mesh ships with 167,529 boundary edges / 4,128 hole loops.
• The bake-time Metal stack (o_voxel.postprocess.to_glb) does its own
  hole-fill on the decimated 200K-face mesh, but it can only close holes
  that survived the decimation step — many of the original decoder holes
  are gone before that point.
• Final GLB renders with see-through cylinders, gear-edge fragmentation,
  and z-sorting artifacts in any glTF viewer (donmccurdy, three.js,
  Blender's importer).

After re-enabling decode-time fill_holes:

• Closes the 4128 hole loops in <1s on the full 1.44M-vertex mesh.
• Final GLB face count: 994,484 vs upstream CUDA's 997,498 (within 0.3%).
• Visual parity with upstream_ref.glb in donmccurdy's viewer.

Changes

1. patches/mps_compat.py: patch_mesh_base()

   • No longer prepends return # Skip — Metal cumesh segfaults... to
     the three methods. Instead:
     • Replaces .cuda() with .to(self.device) — fixes the original
       bare-.cuda() bug for MPS, which the skip masked.
     • Adds an if cumesh is None: return guard so CPU-only builds
       without the Metal stack continue to skip cleanly.
     • Leaves the actual cumesh body in place so it runs normally when
       cumesh is importable.

2. patches/mps_compat.py: patch_pipeline_runtime_fallback() (new)

   • Wraps the decode-time m.fill_holes() call site in
     pipelines/trellis2_image_to_3d.py in a try/except RuntimeError.
     Users on unpatched mtlmesh (Apple7/8 without the atomic-fallback
     fix) get a one-time warnings.warn and degrade gracefully —
     "ship mesh with small holes" rather than crash the whole pipeline.
   • This is the safety net for users who pip-install before the
     upstream mtlmesh PR lands.

3. Calls the new patcher in main().

Validation

Apple M1 Pro 16 GB (Apple7), patched mtlmesh installed:

• cumesh.fill_holes runs in <1s on the 1,438,207-vertex /
  2,988,108-face decoder mesh, closing 4128 boundary loops, adding
  ~26K faces.
• Final GLB: 994,484 faces, extents 1.000 × 0.923 × 0.381 vs
  upstream CUDA's 1.000 × 0.918 × 0.386. Visual parity confirmed
  in donmccurdy's viewer.

Apple M1 Pro 16 GB, unpatched mtlmesh (legacy install simulating a user
who hasn't updated):

• Pipeline completes successfully.
• One-time warning emitted:
  UserWarning: cumesh.fill_holes failed (...); skipping. On M1/M2 update mtlmesh to a build with pedronaugusto/mtlmesh#1 (Apple7/8 atomic fallback).
• Output GLB ships with the original small holes (matches pre-PR
  behavior).

Dependency

• Hard requirement for the quality improvement: Float-atomic + per-face propagation fallback for Apple7/8 GPUs (M1/M2) pedronaugusto/mtlmesh#1
• Soft requirement: this PR works regardless — the try/except just keeps
  the old behavior on unpatched mtlmesh.

Related

• Float-atomic + per-face propagation fallback for Apple7/8 GPUs (M1/M2) pedronaugusto/mtlmesh#1 (Apple7/8 atomic-fallback)
• to_glb: percentile-based alphaMode auto-detection (default OPAQUE-friendly) pedronaugusto/trellis2-apple#1 (alphaMode='OPAQUE' default — separate
  bug fix needed for end-to-end correctness on Apple Silicon)
• bug (or my mistake) #5 (FotisK's original investigation that
  surfaced both fixes)