Enhance FlashAttention and FlexAttention benchmarks with GQA

[mfrancepillois]

FlashAttention and FlexAttention benchmarks currently assess the performance of MHA (Multi-Head Attention).
GQA (Group-Query Attention) is a technique widely used in LLM models (e.g. llama-3.1) to improve performance (by reducing memory usage).
FlashAttention and FlexAttention benchmarks should be improved to evaluate the performance of FA using GQA.

[liangan1]

Need to highlight, for the DeepSeek-v3, the MHA computation will become a MQA with 128 heads and the head_dim=512, the large head dim may put new challenges.