[Dispatch] Fold collapse(expand) unit dims

[IanWood1]

Fold unit dimensions for the pattern collapse(expand) by removing any unnecessary unit dims that are expanded out and subsequently collapsed again. This prevents the collapse/expand ops, which contain unneeded unit dims, from being propagated and introducing unit dims back to linalg ops.

---

Doesn't quite resolve #19263 because there are flow.tensor.bitcast ops that are preventing the unit dims from being completely folded:

%expanded_24 = tensor.expand_shape %214 [[0, 1, 2], [3]] output_shape [4, %21, 1, 128] : tensor<?x128xf16> into tensor<4x?x1x128xf16>
%323 = flow.tensor.bitcast %expanded_24 : tensor<4x?x1x128xf16>{%21} -> tensor<4x?x1x64xcomplex<f16>>{%21}
%collapsed_40 = tensor.collapse_shape %323 [[0], [1, 2], [3]] : tensor<4x?x1x64xcomplex<f16>> into tensor<4x?x64xcomplex<f16>>