Revive DiffMjStep as a lean PyTorch–MuJoCo autograd package

[EladSharony]

Product scope

This PR is intentionally product-only. It contains:

• the installable diffmjstep Python package;
• the native C++ reverse-time VJP implementation;
• three user-facing examples (00_basic_step.py, 03_trajectory_optimization.py, and 04_sensor_loss.py);
• the README (with CPU benchmark figures), license/citation files, and package/release metadata.

The test suite, CPU benchmarks, CI workflow, uv.lock, and development configuration live on the linked revive/validation branch. They remain available for maintainers and reproducibility without inflating the product PR.

Relative to master, the product diff is exactly 26 files, 1,795 additions, and 1,971 deletions: a net deletion of 176 lines.

Retained runtime capabilities

• Public mj_step, mj_rollout, and mj_linearize APIs with explicit batch axes.
• MuJoCo C rollout plus the automatically selected, bounded-thread native C++ reverse-time VJP on Linux and macOS.
• Differentiable sensor losses and standalone linearization.
• Callback-safe replay and isolated, coherent warm-start behavior.

Benchmarks

The README embeds two hand-authored, theme-safe SVG figures (assets/), rendered from measured runs:

• Hero (benchmark.svg): CPU throughput versus a single-threaded pure-Python reference of the same algorithm — 51× faster forward and 13× faster backward at batch 4096 (transitions/s, log scale).
• Per-model (speedup_per_model.svg): the out-of-the-box breakdown at batch 256 across nine dm_control models — forward 11–23× (scales with model simplicity), backward a steady ~5–6×.

Measured on an Intel i9-14900K, MuJoCo 3.10, float64. The baseline is the same mjd_transitionFD reverse-time VJP without the native acceleration, so the speedup isolates the native/parallel backend, not a different gradient method.

Validation evidence

On the product branch:

• uv build completed successfully.
• The wheel was installed into an isolated environment; diffmjstep imported successfully and its package source resolved from the installed wheel environment.
• All three retained examples ran successfully.

On the separate revive/validation branch:

• uv run --frozen pytest -q — 110 passed.
• uv run --frozen ruff check . completed successfully.
• uv build completed successfully.
• The ci run passed on Ubuntu and macOS.