Support fusing Transpose + MatMul where both inputs are transposed

[robertknight]

The decoder of the Whisper example that was recently added spends a large fraction of the time in Transpose operations, especially for larger models. Many of these come from subgraphs that look like:

[Input A] -> Transpose -> MatMul ---> [Output]
[Input B] -> Transpose ------^

Where [Input A] is computed by earlier parts of the graph and [Input B] is a KV-cache tensor from a previous run. This subgraph is inside an If subgraph when using the "merged" decoder, so [Input B] is a captured view.

The graph optimizer currently supports fusing transpose + matmul when only one of the inputs is transposed, but not when both are.