[mlir][Vector] Use create_mask in transfer mask materializations

Currently, the transfer mask is materialized by generating the vector
comparison: [offset + 0, .., offset + length - 1] < [dim, .., dim]

A better alternative is to materialize the transfer mask by using the
operation: `vector.create_mask (dim - offset)`, which will generate
simpler code and compose better with scalable vectors.

Differential Revision: https://reviews.llvm.org/D120487