Add full-horizon waypoint APG training foundation

[yuecideng]

Description

Adds the reusable RL foundation needed to reproduce the full-horizon NMG training contract from neural_motion_generator PR #6.

• Adds scheduled variable-horizon differentiable rollouts with full K-dependent horizons, reciprocal-K objective scaling, first-done masking, gradient accumulation, and exact optimizer-update budgeting.
• Adds overflow-safe per-environment action-adjoint clipping, semantic-mask-aware Welford observation normalization, deterministic lightweight-runtime seeding, and checkpoint support for the new runtime state.
• Extends the lightweight training CLI, evaluation path, public API docs, project context, and focused tests.

The waypoint Transformer model and Franka task implementation are intentionally excluded from this PR. They live together on the pushed downstream branch codex/franka-waypoint-nmg; no PR has been opened for that branch. This branch depends on the Newton integration already present in the target branch and adds no package dependencies.

Type of change

• Enhancement (non-breaking change which improves an existing functionality)
• New feature (non-breaking change which adds functionality)
• Documentation update

Screenshots

Not applicable.

Validation

• black . — 869 files unchanged
• python docs/scripts/check_api_docs.py — 1876/1876 public exports documented
• pytest -q tests/learning tests/docs/test_check_api_docs.py — 77 passed, 3 deselected
• Downstream Franka parity branch compared PR Update default physics mode #6 on CPU and CUDA: joint, Cartesian, and mixed task tensors/one-step gradients had zero max error; a complete 30-step APG update also had zero error for the action trace, loss, gradient norm, every updated actor parameter, and observation-normalizer state

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).
• I have added tests that prove my fix is effective or that my feature works.
• Dependencies have been updated, if applicable (no new dependencies).