Refactor RL advantage estimators and add REINFORCE++ baseline/RLOO support

[listar2000]

Summary

This PR refactors the experimental rLLM advantage estimation flow and adds new estimator options.

Changes

• Refactored advantage estimator interface to return (advantages, returns) per group.
• Added new estimators:
  • reinforce_plus_plus_baseline
  • rloo
• Added rllm/experimental/common/rl_algo.py with grouped GRPO/RLOO advantage helpers.
• Updated trajectory-group advantage collection to:
  • batch computation by role,
  • validate shape/length consistency with strict checks,
  • keep role-level reward/advantage metrics.
• Extended estimator enum + config comments/options in:
  • rllm/experimental/common/config.py
  • rllm/experimental/config/rllm/base.yaml
• Updated Tinker loss-function auto-mapping for new estimators and added a safe default fallback.

Testing

• python -m compileall rllm/experimental/common/advantage.py rllm/experimental/common/config.py rllm/experimental/common/rl_algo.py rllm/trainer/tinker/tinker_policy_trainer.py