Add AMD ROCm (gfx942) support for the image→3D generation stack

[ZJLi2013]

Summary

Enable EmbodiedGen's image→3D generation to run on AMD GPUs (ROCm/HIP), by swapping the
CUDA-only libraries for verified ROCm builds plus two small runtime shims. All changes are
additive (new files under docker/); no existing CUDA code path is modified.

Verified end-to-end on an AMD Instinct MI300X: python -m embodied_gen.models.sam3d
(SAM3D backend, no GPT, no texture-bake) produces outputs/splat.ply (6.5 MB 3D Gaussian
Splat) from the bundled sample_00.jpg.

Changes (all new files)

• docker/install_rocm.sh — one-shot ROCm install: requirements minus CUDA libs, numpy<2
  pin, the ROCm dependency swaps (table below), deploys the two shims as sitecustomize,
  and runs an import smoke (PASS/FAIL map).
• docker/Dockerfile.rocm — full-generation ROCm image (rocm/pytorch:rocm6.4.3...2.6.0)
  that runs install_rocm.sh.
• docker/spconv_rocm_compat.py — converts spconv KRSC checkpoints to the Native layout at
  load time (see Related issue).
• docker/kaolin_stub.py — sitecustomize bypass for the CUDA-only kaolin (used only in
  the texture-backprojection / mesh-IO stage; core geometry path only calls
  kaolin.utils.testing.check_tensor).
• docker/README.rocm.md — user-facing run-through.

CUDA → ROCm dependency map

Upstream (CUDA)	ROCm replacement
spconv-cu120/121	ZJLi2013/spconv_rocm (2.3.8+rocm1, source)
nvdiffrast	ZJLi2013/nvdiffrast@rocm
gsplat	amd_gsplat (pypi.amd.com/rocm-6.4.3; import name stays gsplat)
pytorch3d	ROCm 6.4 / py3.12 prebuilt wheel
flash-attn	FA2-Triton (FLASH_ATTENTION_TRITON_AMD_ENABLE=TRUE at install + runtime)
xformers	not needed — SAM3D attention auto-selects sdpa
numpy (base = 2.x)	pinned <2 (diffusers/transformers requirement)
kaolin (no ROCm wheel)	sitecustomize stub (docker/kaolin_stub.py)
diff-gaussian-rasterization	optional ('inria' GS backend); gsplat is the default

Tested on

• GPU: AMD Instinct MI300X (gfx942)
• ROCm: 6.4.3
• PyTorch: 2.6.0 (+HIP 6.4)
• Docker: rocm/pytorch:rocm6.4.3_ubuntu24.04_py3.12_pytorch_release_2.6.0

Results

• outputs/splat.ply (6.5 MB) from apps/assets/example_image/sample_00.jpg
• Running cost 28.9 s, Max VRAM 9.74 GB; attention on AOTriton SDPA

Notes / scope

• Backward-compatible: only adds files under docker/; CUDA users are unaffected.

• Out of scope (documented gaps, not regressions): texture-backprojection (kaolin is
  CUDA-only and stubbed), GPT quality-checkers (need an API key). Core image→3D
  (segmentation → SAM3D geometry + gaussian + mesh export) runs without them.

• Optional follow-up (happy to include if desired): make the kaolin imports in
  embodied_gen/data/utils.py lazy/optional so the stub isn't needed.

• Depends on / related: spconv KRSC checkpoint loading on ROCm — ZJLi2013/spconv_rocm#<pr>.
  Until merged, docker/spconv_rocm_compat.py provides the equivalent fix consumer-side.

• License: this PR is for study/research purposes only and adds ROCm build/integration
  scripts; it ships no model weights. Any models used (e.g. SAM-3D-Objects, TRELLIS, Kolors,
  SD3.5, etc.) remain governed by their own respective licenses — please refer to each model's
  license before use.