Avoid fully unrolling loops for insert-rotate

After https://github.com/google/heir/pull/587 the box_blur_64x64 test takes ~53 minutes to compile (up from 1.5 minutes) and canonicalizer takes ~50 minutes of that (95%). Long term we're going to need to be able to run our analyses without doing full loop unrolls, and in this case it's primarily to support identifying a target slot for a given op, or perhaps identifying an affine map that converts the extraction index to its target slot that applies to the whole loop, and upgrading insert-rotate to handle loops.