Add latency and body-mass randomization across simulators

[tesslerc]

Summary

• add continuous per-environment action-latency randomization across all simulator backends. Commands are held for the sampled delay, with fractional arrival inside a physics substep represented by a single time-averaged transition. Reset environments receive their first command without stale action history, and invalid delays beyond one control interval are rejected.
• add selective robot body-mass randomization by body name or index, using absolute masses sampled linearly or log-uniformly from configurable buckets. IsaacGym, IsaacLab, and Newton also scale inertia consistently with mass; unsupported backends fail immediately instead of silently ignoring the setting.
• move contact-label smoothing to one vectorized device-side operation while preserving variable-length motion boundaries and replicate-padding semantics.
• prevent evaluation-time autograd state from accumulating, improve MuJoCo keyboard input and viewport recording, and remove redundant IsaacLab viewer rendering work.

The release also includes supporting launcher reliability improvements, documentation, and regression coverage for these behaviors.

Verification

• lint, secret scan, CPU tests, and the full LFS-backed test suite passed
• the projected public tree built successfully, produced a wheel, and passed the release contract
• the projection is a superset of the current public tree
• pretrained-model files and configuration sidecars are byte-identical to public main