Boosting LLM Reasoning: Push the Limits of Few-shot Learning with Reinforced In-Context Pruning