FlashRFCL: Demonstration-Efficient Sparse Reward RL via Reverse-Forward Curriculum Learning in High-Throughput GPU Simulation