[Feature Request] Add Fusion Transformer for WebNN EP Decomposed GQA Node

[Yuhengwe1]

Describe the feature request

We propose adding a fusion transformer to reconstruct Group Query Attention (GQA) nodes that get decomposed during WebNN EP graph processing.

Describe scenario use case

In #23416, the WebNN EP breaks down one GQA node into smaller primitive ops to meet WebNN API constraints which will prevents optimal hardware utilization. Here is the decomposed WebNN subgraph:

Introducing a fusion transformer would reduce the operator count and enable hardware-accelerated attention kernels for WebNN produced models.

/cc @Honry @huningxin

[huningxin]

/cc @fdwr @guschmue @RafaelCintron