[vector][mlir] Canonicalize to shape_cast where possible

[newling]

Discussions suggest that we should use shape_cast as a canonical form of broadcast/transpose/extract where possible (see #138777)

For example these can all be expressed as shape casts:

%0 = vector.broadcast %arg0 : vector<4xi8> to vector<1x1x4xi8>
%1 = vector.transpose %arg1, [1, 0] : vector<2x1xi8> to vector<1x2xi8>
%2 = vector.extract %arg2[0] : vector<4xi8> from vector<1x4xi8>

This PR adds canonicalizes to convert the above 3 examples to shape_casts. It also removes patterns for

shape_cast(transpose) -> shape_cast and
transpose(shape_cast) -> shape_cast

as these foldings are now reachable via the new canonicalization. Specifically

shape_cast(transpose) -> shape_cast(shape_cast) -> shape_cast and
transpose(shape_cast) -> shape_cast(shape_cast) -> shape_cast

IMO the code is overall simplified by this PR. Other notes to reviewers:

• function static bool isOrderPreserving unchanged, just moved closer to usage.
• test movement: tests of shape_cast(broadcast) / shape_cast(transpose) / transpose(shape_cast) removed or moved to new test file.
• PR will be a bit simpler after [mlir][ArmSME] Remove ConvertIllegalShapeCastOpsToTransposes #139706 (scalable vectors) so let's get that in first