feat: align G1 rough-terrain training with MuJoCo Playground

[TATP-233]

Closes #312

This PR implements the G1 rough-terrain training path as a standalone task owner, aligning with MuJoCo Playground G1JoystickRoughTerrain semantics.

What's included

• Independent rough-terrain task owner YAML (selected via task=...)
• Observation / reward / command configuration aligned with UniLab env contract
• Domain randomization and training semantics injected through config
• 1-iteration smoke run validation for the rough-terrain training链路
• Tests for config compose, env obs contract, and training entry
• Documentation on alignment scope and gaps vs. MuJoCo Playground

Checklist

• Independent rough-terrain task owner YAML
• UniLab env contract compatibility (reset/step, obs_groups_spec)
• Config-first reward/command/DR injection
• Smoke run validated
• Tests added at risk boundaries
• Alignment doc included