feat: support Python 3.11 for simulation and training

[wlgys8]

Summary

• Support Python 3.11 for simulation and training packages; keep Unitree hardware runtime and full installations on Python 3.10.
• Add explicit Python selection to the installer, expand CI to Python 3.10/3.11, and update installation documentation.
• Copy callable dataclass configuration defaults per instance for Python 3.11 compatibility.
• Preserve the macOS PyPI Torch fallback through explicit platform resolution partitions in the lockfile.
• Consolidate generic NUMA discovery and placement in motrix_rl.numa; keep FastSAC-specific collector slicing in the provider.
• Dispatch GPU locality queries explicitly by CUDA/HIP runtime and isolate HIP PCI discovery in a subprocess with Torch-wheel library discovery.

Validation

• prek run --all-files: passed.
• Static Python syntax parsing and git diff --check: passed.
• Offline lock consistency check: passed.
• Tests were not run successfully; the local environment lacks pytest and Torch cannot import due to missing libcudnn.so.9. Subsequent work was validated statically as requested.

Limitations

• HIP GPU behavior has not been verified on physical hardware.
• NVIDIA NUMA discovery retains the existing NVML physical-ordinal behavior; CUDA_VISIBLE_DEVICES remapping is not addressed here.

[Copilot]

🔵 Needs a closer look

It broadly alters cross-Python packaging, lockfile resolution, installer shell logic, CI, and a framework-wide @configclass default-copying semantics change, with the author noting the test suite was not run and HIP paths are unverified on hardware.

No open issues 🎉

What changed in this PR

This PR extends MotrixLab to support Python 3.11 for simulation and training while keeping the Unitree hardware runtime and full (--all) installs pinned to Python 3.10. It relaxes requires-python across workspace packages, adds a --python selector to install.sh with an --all/3.11 guard, expands CI to a Python 3.10/3.11 matrix, and updates installation docs in English and Chinese. It also carries two correctness-oriented refactors that 3.11 surfaced: @configclass now deep-copies callable dataclass defaults per instance (3.11 rejects unhashable mutable defaults), and generic NUMA/GPU-locality logic is consolidated into motrix_rl.numa with an explicit cuda/hip runtime dispatch (HIP PCI discovery isolated in a subprocess).

Changes:

• Allow Python 3.11 (requires-python = ">=3.10,<3.12"), add installer/CI/doc support, and exclude Unitree on 3.11 via dependency markers and test importorskip/--ignore.
• Fix @configclass to wrap callable dataclass defaults (e.g. reward/observation term configs) in per-instance deepcopy factories.
• Move gpu_numa_node, spawn_placement, apply_binding, and NVML/HIP discovery into motrix_rl.numa; gpu_numa_node now dispatches explicitly on cuda/hip, keeping only select_collector_cpus in the FastSAC provider.

File	Description
*/​pyproject.toml (11 pkgs)	Relax requires-python to >=3.10,<3.12; keep Unitree marker-gated in test group; add macOS resolution partition.
install.sh	Add validated --python option and --all/3.11 incompatibility guard; thread interpreter choice into uv sync.
.github/​workflows/​tests.yml	Matrix over Python 3.10/3.11; 3.11 path uses direct uv sync and ignores Unitree tests.
motrix_env_core/​.../​config/​decorate.py	Drop callable(value) skip so callable configclass defaults are deep-copied per instance.
motrix_rl/​src/​motrix_rl/​numa.py	Add spawn_placement, apply_binding, NVML/HIP probes, and gpu_numa_node(device_index, runtime).
motrix_rl_builtin/​.../​async_impl/​{numa,topology,train,worker}.py	Reduce provider numa.py to CPU slicing; import shared helpers; pass explicit GPU runtime.
motrix_env_core/​tests/​test_configclass.py, motrix_rl/​tests/​test_{numa,gpu_numa}.py, */​tests/​test_*	Tests for configclass copy semantics, HIP probe isolation/caching, and runtime dispatch; Unitree imports made optional.
README*.md, CONTRIBUTING.md, AGENTS.md, docs/​.../​installation.md, wiki/​.../​fastsac-*.md	Document dual Python support and the NUMA consolidation.

_{Review effort: Balanced}

[Copilot]

🔵 Needs a closer look

It spans build/CI infrastructure, a core @configclass default-copy semantics change affecting all config classes, and platform-specific NUMA/HIP code that cannot be hardware-verified, plus an undocumented Hygon install path, warranting human review.

1 open issue

• PR adds undocumented Hygon DCU GPU profile · New

_{Review effort: Balanced}