[MLIR] Fix tiling for `tensor.unpack` with outer permutations

[MLIR] Fix tiling for `tensor.unpack` with outer permutations

An outer dim permutation requires adjusting the offsets and sizes of the
`tensor.extract_slice` operations generated during tiling. Originally
this was done by computing an inverse permutation of the outer
permutation for both `tensor.pack` and `tensor.unpack`. For packing, the
tiling is applied on interchanged dimensions; thus, it is correct to
compute the inverse. For unpacking, on the other hand, tiling involves
the output tensor that does not have interchanged dimensions, and no
inverse is required.

Reviewed By: hanchung

Differential Revision: https://reviews.llvm.org/D141688