feat: add G1 whole-body tracking task for FastSAC on MuJoCo (#292)

[LeeLeno]

Summary

• Migrate holosoma's G1 WBT task to UniLab FastSAC + MuJoCo
• New env G1MotionTrackingSAC with asymmetric actor-critic (critic gets base_lin_vel)
• Hydra config aligned with holosoma g1-29dof-wbt-fast-sac hyperparameters
• Verified 2.3x throughput improvement over holosoma (IsaacSim) on same hardware with AMP enabled
  also:
• Add G1MotionTrackingSACEnv — a thin SAC wrapper over G1MotionTrackingEnv that appends base_lin_vel to critic observations (asymmetric actor-critic, aligned with holosoma g1-29dof-wbt-fast-sac)
• Add off-policy task config conf/offpolicy/task/sac/g1_sac_wbt/mujoco.yaml with holosoma-aligned hyperparameters (gamma=0.99, tau=0.05, num_atoms=501, target_entropy_ratio=0.5)
• Add motion asset sub3_largebox_003.npz for WBT training
• Fix MuJoCo render camera: support cam_distance parameter passthrough and adjust default lookat height to 0.75
• Tune global off-policy defaults: max_grad_norm 0→10, play_steps 200→800
• Update docs (support matrix & motion-tracking guide) with FastSAC WBT instructions

Usage

# Default training (MuJoCo, 2048 envs)
uv run scripts/train_offpolicy.py algo=sac task=sac/g1_sac_wbt/mujoco

# With AMP for faster training on CUDA
uv run scripts/train_offpolicy.py algo=sac task=sac/g1_sac_wbt/mujoco training.use_amp=true

# Custom env count and iterations
uv run scripts/train_offpolicy.py algo=sac task=sac/g1_sac_wbt/mujoco \
    algo.num_envs=4096 algo.max_iterations=10000 training.use_amp=true

# With wandb logging
uv run scripts/train_offpolicy.py algo=sac task=sac/g1_sac_wbt/mujoco \
    training.use_amp=true training.logger=wandb

# Replay with checkpoint (after training)
uv run scripts/train_offpolicy.py algo=sac task=sac/g1_sac_wbt/mujoco \
    training.play_only=true \
    training.checkpoint=<path_to_checkpoint>

Test plan
 make check (ruff + mypy + pyright)
 make test (non-slow tests, includes auto-discovered config validation)
 Train smoke test: uv run scripts/train_offpolicy.py algo=sac task=sac/g1_sac_wbt/mujoco algo.num_envs=64 algo.max_iterations=50 algo.batch_size=512
 Existing tasks unaffected: uv run scripts/train_offpolicy.py algo=sac task=sac/g1_walk_flat/mujoco algo.num_envs=64 algo.max_iterations=10