feat(appo): restore Go2 Motrix owner on tensor runtime

[TATP-233]

Summary

• restore appo/go2_joystick_flat/motrix as a configured tensor-runtime owner
• register Go2JoystickFlat for Motrix
• upgrade the optional/dev unilab-rl dependency to 1.4.8
• follow the upstream SAC package rename (fast_sac → sac, FastSACLearner → SACLearner)
• route APPO through the same tensor EnvFactory as off-policy training, with the tensor NaN-guard factory
• regenerate bilingual support-matrix evidence and update Motrix scope docs
• add a bounded Motrix APPO slow smoke that asserts model_1.pt

Dependency changes

unilab-rl 1.4.8 includes:

• tensor-native APPO collector, rollout IPC, and staging (upstream chore: set up github collaboration workflow #84)
• Torch action/reset/state env contract
• Gaussian std metrics derived directly from distribution parameters
• CPU-authoritative env + CUDA collector bootstrap correction fix
• SAC package/learner rename with no compatibility layer

This PR removes the need for a consumer-side APPO NumPy adapter or custom MLPModel subclass.

Validation

Local gates:

make check
# passed

make test
# 1893 passed, 28 skipped, 552 deselected

make test-all
# passed, including benchmark smoke 35/35 modules and 36/36 entrypoints

Focused runtime tests:

uv run pytest -q \
  tests/scripts/test_train_script_configs.py \
  tests/algos/test_appo_runner.py \
  tests/algos/test_offpolicy_double_buffer_runner.py \
  tests/algos/test_offpolicy_dp_sync.py \
  -m 'slow or not slow'
# 70 passed, 2 skipped

Bounded real training through the public CLI route:

MuJoCo:
  status=completed
  completed_iterations=1
  total_env_steps=40
  checkpoint=model_1.pt

Motrix:
  status=completed
  completed_iterations=1
  total_env_steps=16
  checkpoint=model_1.pt

Both runs use the published unilab-rl==1.4.8 package.