Efficient encoding of memref's layout mapping

[aqjune]

Sometimes simply encoding the mapping (i, j, k) -> blk offset using UF is more efficient than writing its full expression.

Previously we had this option but is removed in 168f52a because the option isn't used.