Determine how to hoist layout conversions through bicyclic matmul

A continuation of https://github.com/google/heir/issues/1376

Comment from @asraa on https://github.com/google/heir/pull/2360:

> For some layouts there should be a hoister. if there's a result layout that happens to be a rotation of the bicyclic encoding, then I think that the inputs could also be rotated. but agree to file and we could work it out on paper.
but i also don't think that would happen organically, and it's more likely that we would be hoisting a row-major layout through a bicyclic encoding.

Comment from @jkun:

> It's not clear how to hoist a layout conversion through the bicyclic kernel in a meaningful way. So currently we have no hoister. Maybe it's worth just putting in a "trivial" hoister that results in new layout conversions for each operand back to bicyclic, since they may fold with the hoisted layout conversion to be simpler... but it seems unlikely.