Add Newton physics backend support

[yuecideng]

Description

This PR makes Newton a first-class physics backend in EmbodiChain while preserving the existing default DexSim path. It also standardizes scene construction around DexSim Spawn and adds an end-to-end differentiable execution path for analytic policy gradient workloads.

The motivation is to keep backend-specific lifecycle, state access, spawning, and differentiation behind explicit contracts instead of scattering Newton conditionals through simulation objects and environments.

Issue reference: N/A.

What changed

Physics backend and lifecycle

• Adds typed DefaultPhysicsCfg and NewtonPhysicsCfg configuration, including Newton solver, collision-pipeline, device, substep, CUDA graph, and gradient settings.
• Introduces the PhysicsBackend abstraction with concrete default and Newton implementations selected from physics_cfg.
• Moves Newton ownership to the world/Spawn lifecycle and exposes SimulationManager.prepare(), backend capabilities, scene context, and differentiable-runtime delegators.
• Handles multi-environment spawning, rebuild/rebinding, multi-instance isolation, and deterministic cleanup through explicit runtime contracts.

Spawn and object model

• Adds embodichain.lab.sim.spawn to translate EmbodiChain configs and USD sources into DexSim Spawn descriptors.
• Adds backend-neutral rigid-body and articulation views for default, Newton, and Spawn-backed access.
• Supports Newton rigid objects, rigid-object groups, articulations, and robots, including batched pose/state access, physical-property mutation, joint control, collision filtering, and multi-env cloning.
• Adds explicit asset-physics preserve / overlay behavior and typed backend-specific rigid, collision, material, articulation-root, and joint-drive properties.
• Introduces a unified deformable-object API with volume/soft and surface/cloth specializations while retaining compatibility aliases.

Differentiable execution

• Adds a Warp tape to PyTorch autograd bridge in embodichain.lab.sim.diff.
• Adds DifferentiableEmbodiedEnv with explicit dynamics and kinematics execution modes, manager-owned Newton trajectories, deferred terminal resets, and safe forward/backward cleanup.
• Adds the FrankaReachApg-v0 reference task to exercise differentiable stepping and analytic policy gradient behavior end to end.

Compatibility, examples, and documentation

• Preserves legacy simulation configuration accessors and the default-backend behavior covered by the compatibility tests.
• Updates environments, sensors, planners, tutorials, benchmarks, and examples to use the explicit prepare/Spawn lifecycle.
• Updates API reference pages, simulation guides, design documents, and agent context for the backend, Spawn, object, and differentiable-environment contracts.
• Adds focused coverage for configuration parsing, backend parity, Spawn descriptors, rigid/articulation behavior, lifecycle, multi-env execution, and gradients.

Backend capability summary

Capability	Default	Newton
Rigid objects	Yes	Yes
Articulations / robots	Yes	Yes
Rigid-object groups	Yes	Yes
Volume deformables / soft bodies	Yes	No
Surface deformables / cloth	Yes	No
Disable manual update	Yes	No
Differentiable autograd bridge	No	Yes, with gradient mode

Known limitations

• Newton volume and surface deformables remain unsupported and fail fast through backend capability checks.
• Newton cannot disable manual update.
• Differentiable dynamics require NewtonPhysicsCfg(requires_grad=True) with the semi_implicit solver.
• The Franka APG example intentionally uses the explicit kinematics route because the semi-implicit target-position path does not currently propagate gradients to body poses.
• Newton callers should use SimulationManager.spawn_result and Spawn batch APIs; the removed independently owned NewtonManager / compatibility PhysicsScene ownership path is not exposed.

Dependencies

No new dependency is added by this branch. The implementation requires the Spawn/Newton APIs provided by the existing dexsim_engine==0.4.3 dependency.

Suggested review order

1. embodichain/lab/sim/cfg.py, physics/, and sim_manager.py
2. embodichain/lab/sim/spawn/ and objects/backends/
3. Rigid object, articulation, robot, and deformable-object integrations
4. embodichain/lab/sim/diff/, DifferentiableEmbodiedEnv, and the Franka APG task
5. Tests, examples, and documentation

Type of change

• Bug fix (non-breaking change which fixes an issue)
• Enhancement (non-breaking change which improves existing functionality)
• New feature (non-breaking change which adds functionality)
• Breaking change (existing functionality will not work without user modification)
• Documentation update

Screenshots

Not applicable.

Validation

Focused post-main merge validation:

• Formatting checks passed with Black 26.3.1.
• Public API documentation gate passed: 1,798 documented exports, 0 missing.
• Preview asset, gravity settler, tutorial utilities, and articulation checks: 57 passed, 1 skipped.
• Semantic-skill tutorial checks: 12 passed.
• Rigid-object CPU and Newton checks: 6 passed.
• Agent-context registry checks: 8 passed.
• API-documentation checker tests: 8 passed.
• Sphinx dummy build completed successfully; optional-import and existing documentation warnings remain.
• GitHub CI is still running for the latest head.

Checklist

• I have run the Black formatting checks.
• I have made corresponding changes to the documentation.
• Public API changes are reflected in the API docs.
• I have added tests for the affected backend, Spawn, object, and differentiable paths.
• Dependencies have been reviewed; no new dependency was added.