Audio Question Answering with GRPO-Based Fine-Tuning and Calibrated Segment-Level Predictions