[`LinalgToXeGPU`] Support conversion for `linalg.matmul` with `transpose_b`

[dchigarev]

This can be implemented using xegpu.load_nd %x {transpose = {1, 0}} for B tiles.

We should support both patterns:

1. Explicit transpose in op linalg.matmul_transpose_b.
2. Transpose before matmul (this is how MLIR from OV will look like):

   %b_tr = linalg.transpose %b ...
   %res = linalg.matmul %a, %b, ...

This functionality is required for OV integration (#207).