Fix RL joint isolation and expose batched simulation controls

[acrlw]

Description

This PR fixes active-joint list sharing between environments and exposes the root-velocity and scene-capacity controls needed for batched locomotion training.

Changes

• Isolate active-joint lists. Creating a second environment through control_parts can change the first environment's joint list from [0, 2] to [0, 2, 0, 2]. EmbodiedEnv._setup_robot() now builds a fresh instance list on each setup and copies explicitly configured active_joint_ids.
• Expose root-velocity writes. Push events calling Robot.set_root_velocity() previously raised AttributeError. Add Articulation.set_root_velocity(velocity, env_ids=None), inherited by Robot, and the corresponding view interface. The input is an (N, 6) tensor containing world-frame linear velocity followed by angular velocity. SceneArticulationView validates the shape, selects one batch for both writes, preserves row order, and leaves unselected environments unchanged.
• Expose material reuse. Add DefaultPhysicsCfg.cache_material, defaulting to False, and forward it through to_dexsim_args(). Training configurations can enable DexSim's existing material reuse with physics_config: {cache_material: true} for the default physics backend.
• Forward scene-node capacity. Add optional SimulationManagerCfg.scene_node_capacity and forward the training configuration through config_to_cfg() to WorldConfig.scene_node_capacity. None preserves the native default; an explicit value such as 262144 reaches World creation.

Public API usage, configuration documentation, and the affected project context are updated.

Dependencies: no new dependencies. The changes use existing DexSim batch and configuration APIs.

Validation

Validated on the combined branch based on main@7bc7271c, using Python 3.11 and a local RTX 5090 D v2:

Check	Result
Object/view, joint-isolation, material-cache, and node-capacity tests	47 passed
Existing simulation-configuration tests	131 passed
Native root-velocity writes	CPU and GPU runs: three instances each, selection [2, 0], zero maximum readback error, and finite values after three physics steps
Project context	Registry check passed; 23 tests passed
API documentation coverage	2,069 exports; no missing entries
Formatting	Black 26.3.1: all 1,042 Python files passed; git diff --check passed

Pending on this branch: 2,000-update training runs and a full Sphinx build.

Type of change

• Bug fix (non-breaking change which fixes an issue)
• Enhancement (non-breaking change which improves an existing functionality)
• Documentation update

Screenshots

Not applicable; this PR changes simulation APIs and configuration.

Checklist

• I have run the black . command to format the code base.
• I have made corresponding changes to the documentation
• Public API changes are reflected in the API docs (python docs/scripts/check_api_docs.py), if applicable
• I have added tests that prove my fix is effective or that my feature works
• Dependencies have been updated, if applicable.