Add grouped scene randomization with saved layout preview

[MuziWong]

Description

Add grouped scene randomization for existing Scene Engine exports, building on the shared procedural generator and AABB layout solver from #761.

• Group each direct table child with its descendants and cache source geometry and relative poses in a validated baseline YAML.
• Accept a versioned YAML policy with seed, candidate_count, max_attempts, and collision_margin_m. Generate fixed-orientation XY layouts within the table bounds while preserving each group's internal poses. Return partial results with budget_exhausted when the attempt budget runs out.
• Preserve exported collision settings in returned scene fragments: the default 32-hull decomposition, explicit single hulls, and custom VHACD hull budgets.
• Save each run's resolved policy, candidate JSON files, and completion manifest. Reload and cycle the original scene and saved layouts through the existing preview, loading assets once and applying poses without advancing physics.
• Document all 12 new public APIs, policy and playback examples, and GenSim context routing.

Generation: embodichain scene-randomization --output_root <workspace> --config <policy.yaml>

Playback: python -m embodichain.gen_sim.scene_engine.randomization.loading --layout-dir <run>/layouts --interval 3

The initial scope is a single table with rigid GLB objects and explicit support relationships. Layout acceptance checks group AABB bounds and clearances; physical stability and task feasibility require separate validation. No new dependencies.

Type of change

• New feature (non-breaking change which adds functionality)
• Bug fix (preserves collision configuration when consuming generated fragments)
• Documentation update

Preview

Left: fixed input image. Right: native DexSim playback of 20 randomized layouts, shown at 2× speed (about 10 seconds per loop). Assets are loaded once; each support group and its members move together through pose updates without advancing physics.

Validation

• black . using Black 26.3.1: passed; git diff --check: passed.
• Scene randomization and core/export tests: 209 passed, comprising 204 CPU-selected cases and 5 mock-only preview cases run without global simulation initialization.
• The three collision-configuration regression cases fail before the fix and pass afterward, covering omitted/default, explicit single-hull, and custom decomposition budgets.
• tests/test_agent_context_map.py, tests/test_agent_context_tools.py, and tests/docs/test_check_api_docs.py: 49 passed.
• python docs/scripts/check_api_docs.py: passed, zero missing exports.
• context.py check: passed; CLI, Chinese grouped-scene queries, and public symbols route to GenSim while domain randomization retains its existing owner.
• The new API documentation sections pass a Sphinx dummy build with warnings treated as errors. The full documentation dummy build succeeds with existing warnings elsewhere.
• Local runtime tests used the installed DexSim development build with its existing JIT import-path compatibility shim. These automated checks did not launch a live simulator, GPU, or renderer; the preview above was recorded separately with native DexSim.

Checklist

• I have run the black . command to format the code base.
• I reviewed affected documentation and agent context, updated it where needed, or explained why no update was needed.
• Public API changes are reflected in the API docs (python docs/scripts/check_api_docs.py).
• I have added tests that prove my fix is effective or that my feature works.
• Dependencies have been updated, if applicable (no dependency changes required).