Gemm/Transpose fusion - additional pattern coverage

[ytaous]

Description: In the final optimized training graph of ULRv5 model, found one Transpose/Gemm pattern not fused properly. Thus adding additional logic to cover the case when a Transpose output is consumed by multiple Gemm nodes.

The new pattern only consider the case where all consumers of a Transpose are Gemm nodes. If one of the consumer node is not Gemm, we'll skip the fusion.

Example:
Before:

Expected results after transformation:

Motivation and Context

• Why is this change required? What problem does it solve?
• If it fixes an open issue, please link to the issue here.