Revise the baby step giant step ratio when zero diagonals are present

https://github.com/google/heir/pull/2449 skips and zero diagonals present in the filter matrix.

Instead, we may actually want to refactor our implementBabyStepGiantStep to just take in the non-zero diagonals. The original diagonalized tensor matrix (say, tensor<1024x1024>) can be extracted to just the rows that are non-zero (maybe tensor<241x1024>) and these remaining rows should be used in the baby-step giant step calculation.

This would require extracting the right rows out of the original matrix, and they may not always be contiguous so extract_slice may not work.

As a result, we would lower the # of rotation automorphisms and key switches