v0.3.0

Added

• Full Reinforcement Learning pipeline for training custom backtracking policies (backtracking-llm-train-rl).
• RlPolicyOperator for executing trained PPO policies during inference.
• Gymnasium environment with LLM-as-a-Judge scoring and intermediate reward shaping.
• GenerationSession class for fine-grained, stateful control over the generation loop.

Fixed

• Fixed critical state leakage in CLI chat sessions where operators retained history between turns.
• Fixed "amnesia" bug in NGramOverlap where backtracking erased too much history.
• Fixed greedy decoding to be truly deterministic when temperature is set to 0.