Fix cloop and warnings

[Arul-Sujith]

This PR implements a fix for the cloop! macro's step_by variant as discussed in issue #1504.

Changes

• The step_by variant of the cloop! macro in both ml-kem and ml-dsa was mathematically incorrect for cases where $start != 0 or $end % $step != 0. This replaces the flawed implementation with a rigorously bounded for loop 0..(($end) / $step + 1) that computes the offset internally. This maintains Eurydice's ability to extract cleanly to C (unlike loop { ... } approaches) while preventing loop bound overshoots and $start scaling bugs.