[DML EP] Contrib Op: FusedMatMul

[sumitsays]

Description:

• Adds the Contrib operator kernel FusedMatMul to the DML EP.
• Changes the name (prefix) of the operator fusions performed by DMLGraphTransformer to from Fused* to DmlFused*. It is needed to avoid the name overlapping between ORT Contrib operator FusedMatMul and DML fused operator FusedMatMul (now this is DmlFusedMatMul)
• According to ORT contrib operator FusedMatMul shape inference, the sequence is:
  • Honor the attributes trans and transBatch by manipulating the input shapes and generating the appropriate strides.
    • trans: transpose the last 2 dimensions.
    • transBatch: shift the batch dimensions (all dimensions expect the last 2 dimensions) by 1 and move the 2nd-last dimension to the place of 1st dimension.
    • Between the transBatch and trans, apply transBatch first and then trans
  • Per the numpy MatMul behavior, broadcast the shapes of the input tensors and accordingly manipulate their strides (Note: At this point strides may have already been calculated, if attribute trans or transBatch is present, and those strides should not be overridden).
• This kernel is a special one because it does not care about actual input shape (comes from the .onnx file) while creating the DmlOperator::TensorDesc because as mentioned above actual input shapes need to be manipulated to acknowledge the 2 types of transposes. That's why I added a new way to initialize a DMLOperator via DmlOperator::InitializeWithShape
• winml_test_model.exe successfully passes, which verifies nothing breaks because Dml fused operator re-naming.

• OnnxBackendTest also successfully passes:

Motivation and Context

• Why is this change required? What problem does it solve?
  This contrib operator is used by transformer-based models.

[fdwr]

Needs some robustness checks for span copies.

[fdwr]

🎉