Add deterministic trajectory lookahead

[fei-yang-wu]

What changed

• add a dedicated reset generator to ParallelTrajectoryManager
• expose cursor-only lookahead APIs that compute exact replay indices without reading storage
• allow SONIC reset sampling from a caller-owned probability snapshot
• thread the dedicated generator through random, adaptive, and SONIC reset sampling
• add coverage for deterministic RNG independence, cursor lookahead, validation, and snapshot behavior

Why

IsaacLab-Imitation now stages the exact next reference batch asynchronously while physics runs. That requires predicting replay rows without performing the storage read, and reset prediction must not race the global RNG or later SONIC failure updates.

These APIs keep trajectory bounds and local-to-global index semantics owned by ParallelTrajectoryManager while making deterministic asynchronous consumers possible.

Validation

• pixi run pytest -q ImitationLearningTools/tests/datasets/test_parallel_trajectory_manager.py ImitationLearningTools/tests/datasets/test_reset_sampling.py — 29 passed, 1 skipped
• pixi run test-rlopt — 121 passed
• IsaacLab-Imitation extension suite — 152 passed