[codex] implement AdaVLA-RL adaptive compute

[scout-123-china]

Summary

• Add AdaVLA-RL adaptive compute policy, learned token scoring, compute-cost accounting, and PPO helpers.
• Integrate dynamic visual-token budgets, language-layer depth control, and optional fast-path inference into the OpenVLA HF model.
• Add offline and LIBERO online RL training scripts plus evaluation/deployment loading paths for compute-policy checkpoints.

Validation

• python -m ruff check --select F,E9,B008 prismatic/models/adaptive_compute.py prismatic/extern/hf/modeling_prismatic.py vla-scripts/finetune.py vla-scripts/train_adavla_rl.py vla-scripts/train_adavla_rl_libero.py experiments/robot/openvla_utils.py experiments/robot/robot_utils.py experiments/robot/libero/run_libero_eval.py vla-scripts/deploy.py
• python -m compileall prismatic experiments vla-scripts

Notes

• Direct push to moojink/openvla-oft was denied for the current account, so this PR is opened from the fork scout-123-china/openvla-oft.