rotate-and-reduce: apply for sums of contiguous regions exceeding a threshold

Add a pass option describing some cost function for which the rotate-and-reduce optimization is worth applying to a partial contiguous region of a tensor.

As @asraa mentioned in https://github.com/google/heir/pull/575, we could also try more exotic patterns like summing reductions that are not contiguous but well-strided.

Cf. https://github.com/google/heir/pull/517#discussion_r1520698063