Add Apple Silicon (MPS) support with bf16 autocast

[andeplane]

Summary

Enables demo.py on Apple Silicon GPUs (MPS backend), tested on an M4 / 48 GB with lingbot-map-long.pt on the bundled example/courthouse scene (--use_sdpa, since FlashInfer is CUDA-only).

• Select mps when CUDA is unavailable; run bf16 under a device-aware autocast context
• Make the fp32 guard around the camera/depth/point heads device-aware (autocast('cuda', enabled=False) does not disable autocast on other device types, so heads would silently run bf16 on MPS)
• Downcast precomputed RoPE frequencies from complex128 to complex64 when moving to MPS (MPS has no float64); the fp64 precompute on CPU is unchanged
• Fix matplotlib>=3.9 compat in the viewer (cm.get_cmap was removed in 3.9)

CUDA behavior is unchanged in all paths.

Related PRs

• Add Mac MPS accelerated inference support #17 covers similar ground (MPS device selection with a --device flag, auto-SDPA, the same RoPE complex64 downcast) but runs MPS in fp32 without autocast. This PR runs bf16 on MPS — matching the CUDA inference dtype at roughly half the activation/KV memory — which is what required the device-aware autocast and fp32-head-guard changes above. Happy to rebase onto Add Mac MPS accelerated inference support #17's --device UX if that PR is preferred as the base.
• Add CPU-first inference with automatic SDPA fallback #85 adds CPU-first inference with a utils/device.py device/dtype resolution layer and soft FlashInfer→SDPA fallback; if it lands first, the MPS branch here would slot naturally into those helpers.
• Add Mac metal MLX inference #50 is a native MLX reimplementation — likely the performance ceiling on Apple Silicon, but a separate stack; this PR is the minimal torch-native path using the existing demo.py flags.

Testing

• example/courthouse (286 frames, streaming, --mask_sky --use_sdpa) on M4 (macOS 26, torch 2.13): full run in progress at time of writing (~1.5 it/s early, slowing as the KV cache grows); will confirm completion + viewer output in a comment
• 12-frame smoke test end-to-end through inference, post-processing, and the viser viewer

🤖 Generated with Claude Code