[MLIR][Tensor] Add Destination style RewritePattern for DimOp. (#65780)

This enables canonicalization to fold away unnecessary tensor.dim ops
which in turn enables folding away of other operations, as can be seen
in conv_tensors_dynamic where affine.min operations were folded away.