Fix MPS tensor layout for WiLoR inference

[lyonsno]

Summary

This fixes two non-contiguous tensor boundaries that break or threaten PyTorch MPS inference on Apple Silicon:

• make the cropped ViT/backbone input contiguous after x[:, :, :, 32:-32]
• make refinement image features contiguous before the first convolution in both deconv variants

The patch is intentionally narrow: it does not change model weights, detector behavior, output schema, or CPU semantics.

Verification

Base checked on 2026-05-22: warmshao/WiLoR-mini@ebec42f94c389070cdd7dda6fd1bf0b4a659c960. Open PR #21 was also checked (alex-bene:deps-fix-focal-input-det-conf@cf8f582223142a57f2f486cd7b2fc0545663473d); that PR covers dependency/API work and does not include this MPS tensor-layout repair.

Local test command:

.venv/bin/python -m pytest tests/test_mps_layout.py -q

Current result on PR head e6f1ceee111b3792303c32db8168f4485eb807fa:

6 passed in 0.55s

The layout tests now run on CPU in ordinary CI and additionally on MPS when available. A fail-first check against base with the current test file failed on the intended non-contiguous tensor assertions for the backbone crop and both refinement deconv variants.

Saved-image smoke on Mac MPS, using assets/img.png, float32, predict_with_bboxes, and cached WiLoR-mini model artifacts:

{"torch":"2.5.0","mps_built":true,"mps_available":true}
{"init_seconds":1.655}
{"elapsed_seconds":6.144,"num_outputs":1,"pred_vertices_shape":[1,778,3],"pred_keypoints_3d_shape":[1,21,3],"pred_cam_t_full_shape":[1,3],"device_route":"mps","dtype":"float32"}

CPU smoke with the same saved image/cache also passed:

{"torch":"2.5.0","init_seconds":1.405,"elapsed_seconds":0.298,"num_outputs":1,"pred_vertices_shape":[1,778,3],"pred_keypoints_3d_shape":[1,21,3],"pred_cam_t_full_shape":[1,3],"device_route":"cpu","dtype":"float32"}

Dependency facts from the smoke venv:

{"python":"3.10.19","torch":"2.5.0","torchvision":"0.20.0","ultralytics":"8.1.34","chumpy":"0.71","dill":"0.4.1","mps_built":true,"mps_available":true}

One dependency note: the detector checkpoint load required dill; it was not installed by the current requirements, and Ultralytics' auto-install fallback failed in a uv-created venv because pip was not present. I installed dill locally to continue the MPS smoke, but did not add that dependency here to keep this PR scoped to the tensor-layout repair.