fix(training): keep MuJoCo collector tensors on CPU by default

[TATP-233]

Summary

• Added training.collector_tensor_device: cpu | cuda to the SAC, FlashSAC, PPO, and APPO training owners, with cpu as the default.
• Routed that explicit choice into the existing Manager owner field (manager_torch_device). An unindexed cuda request is resolved by the environment to the current rank-local Torch CUDA ordinal.
• Prevents automatic learner CUDA selection from implicitly moving a MuJoCo HOST_BRIDGE collector/Manager carrier to CUDA. Non-MuJoCo CUDA requests fail closed; DEVICE_RESIDENT backends continue to own their GPU placement.
• Updated ADR-0012 and the maintained English/Chinese installation and support-matrix docs.

Linked Work

• Issue: Fix MuJoCo collector tensor device placement #2100
• Investigation: Bug: SAC G1 MuJoCo throughput regressed after automatic HOST_BRIDGE Manager CUDA placement #2096
• Governing ADR: docs/sphinx/source/adr/ADR-0012-sole-tensor-manager-and-scoped-backends.md
• Parent roadmap: not applicable
• Base branch: main

Validation

• make test-all passed on the final local head before this PR was created
• Additional task-specific validation listed below

Commands actually run:

uv run pytest tests/base/backend/test_process_device.py
# 32 passed

uv run pytest tests/envs/test_manager_based_rl_env.py -k 'host_bridge or manager_tensor'
# 6 passed, 80 deselected

uv run pytest tests/scripts/test_train_scripts.py -k 'collector_tensor_device or host_bridge'
# Exit code 5: 137 collected, 137 deselected. This file is entirely marked slow by existing config.

uv run pytest --override-ini "addopts=--tb=short" tests/scripts/test_train_scripts.py -k 'collector_tensor_device or host_bridge'
# 7 passed, 130 deselected

make check
# Passed: ruff format/check, mypy, pyright, and focused test lint

make test
# 1857 passed, 43 skipped, 495 deselected

make test-all
# Passed, including 1857 passed / 43 skipped and benchmark import smoke 34/34 module-mode and 35/35 script-mode

Complete default SAC validation:

uv run train --algo sac --task g1_walk_flat --sim mujoco \
  algo.max_iterations=160 training.no_play=true training.export_onnx=false \
  training.nan_guard.enabled=false training.logger=tensorboard

Result:

• Completed 160/160 iterations.
• run_summary.json tail FPS: 93,873.5; TensorBoard last-80 median FPS: 97,666.7.
• TensorBoard last-80 median apply_action: 0.279 ms.
• Manifest: collector_tensor_native=false, env_public_device=cpu, CPU inference transport, and learner_device=cuda:0.
• Local artifact: logs/sac/G1WalkFlat/2026-10-09_13-49-45_mujoco/run_summary.json.

Explicit CUDA smoke:

uv run train --algo sac --task g1_walk_flat --sim mujoco \
  training.collector_tensor_device=cuda algo.max_iterations=10 \
  training.no_play=true training.export_onnx=false \
  training.nan_guard.enabled=false training.logger=tensorboard

Result: completed 10/10 iterations with collector_tensor_native=true, env_public_device=cuda:0, and CUDA inference transport.

Remote CI route:

• Base main: current-head CI passed at commit 5290ed822935bf2090d816cd899789b7531e4147.
  • CI: https://github.com/Motphys/UniLab/actions/runs/37893507540 — ruff-format, ruff-lint, mypy, pyright, benchmark-smoke, and test (ubuntu-slim) passed.
  • Docs: https://github.com/Motphys/UniLab/actions/runs/37893507594 — Build Sphinx passed; docs deploy is expected to skip.

Impact

• Backend impact: mujoco
• Platform impact: Linux CUDA training behavior; docs also cover ROCm's CUDA Torch namespace
• Training effect expected: yes — MuJoCo defaults to CPU Manager/collector carriers while retaining an explicit CUDA opt-in

Artifacts

• W&B: none
• benchmark result: local run_summary.json files cited above
• video / screenshot: not applicable
• ONNX / checkpoint: local smoke checkpoints were disabled or incidental and are not attached

Checklist

• Added or updated tests where needed
• Updated docs if behavior or workflow changed
• Linked the driving issue
• Noted any follow-up work explicitly

Follow-up: none required by this scoped change. Motrix/Drake/Superdex remain intentionally outside the #2100 backend boundary.