feat: Native LeRobot integration via make_env() + DDS

[MiaoDX]

Context

Sub-issue of #74. This is the "final form" — using LeRobot's official make_env() factory with full DDS communication, replacing our standalone LeRobotG1Env.

Current state (SDK-free path, working)

# Our current approach: load model directly, bypass DDS
env = LeRobotG1Env(model_path=xml_path, num_motors=29)
ctrl = GrootLocomotionController()  # our ONNX wrapper

Target state (native LeRobot)

# Official LeRobot approach: full digital twin via DDS
from lerobot.envs.factory import make_env
env = make_env("lerobot/unitree-g1-mujoco", trust_remote_code=True)
# LeRobot's own controller handles locomotion via DDS internally

Blockers

• SDK compatibility resolved (see sub-issue: unitree_sdk2_python_uv)
• cyclonedds installs on Python 3.12+ (LeRobot requires >=3.12)
• Verify make_env() returns an env with render_camera() or only render()

Tasks

• Create examples/lerobot_g1_native.py using make_env()
• Wrap with RobotHarnessWrapper — validate Gymnasium API compatibility
• Document full setup (conda env, SDK install, Python version)
• Test in CI (at least one Python version)
• Compare with SDK-free approach: rendering quality, camera support, control fidelity

When to do this

After the SDK fork (#81) is stable and tested. The SDK-free path (#78, #79, #80) is sufficient for now.