Enhancing Performance and Scalability of Large-Scale Recommendation Systems with Jagged Flash Attention