Support Grouped Query Attention (GQA) and update flash attention options

[janpfeifer]

This PR implements Grouped Query Attention (GQA) in the XLA backend and updates the custom-call signature to align with the new compute interface:

• Removed direct mask parameter and moved it into ScaledDotProductAttentionConfig.
• Allowed numHeads to be a divisible multiple of numKVHeads (GQA/MQA).
• Implemented GQA key/value head broadcasting prior to invoking the cuDNN flash-attention custom call.
• Implemented GQA gradient reduction (reduce-sum) for dKey and dValue back to numKVHeads in the VJP backward pass.
• Updated dispatcher and integration tests to match the new signature.