ML-KEM: Try to clean up the NTT implementations

[marsella]

Does not quite finish #147.

This addresses most of the clean-up issues in the linked issue, except for cleaning up the naive NTT version described in the spec (see Algorithms 9 and 10). That has a dense nested loop structure whose bounds depend on each other's variables and that iteratively updates the state of the f variable. We have a version that works but doesn't clearly map onto the spec, and I spent a day failing to find a more obviously-matched way to implement it.
Update: I made a recursive version of Algorithm 9 which may or may not be better. I solicit opinions from reviewers on whether they like it or not.

Everything else should be addressed in a reasonable way. Looking for general gold-standard quality review.

[marsella]

I revisited this today and have not gained any insight in the interim week. The innermost loop of NTT accesses and updates 2 * len elements of the loop, where len is variable based on previous loop iterations. I was hoping I could try to fold across the inner loop, but ran into several problems:

• if the iterator j is a variable, I don't think there's a way to update a single element in the output list (everything I can imagine uses take and drop to preserve the unchanged elements, which would require j to be a type)
• There's no way to "build up" the output list as we go because the fold function has to be the same for every iteration, which means it can't be parameterized by an increasing length parameter.

I keep ending up in the following train of thought:

• If I want to update elements of the list one-by-one, the index of them has to be a type
• If the index is a type I'm going to have to use recursion to update it as we iterate

There might be a somewhat cleaner way to implement the recursive loops than what I did in e742abb and 0b355c3, but I think that's going to have to be the basic structure, which means it's going to be hard to read. Unless we can find a different NIST-standardized NTT description that's stated to be equivalent to this one, and can implement that instead.

Anyway, for now I propose we take the clean-up commits through b56844e and I'll spin out a separate issue to implement NTT itself more closely to the spec.

[marsella]

I reverted deleted my attempt at a recursive version and made #163 to handle fixing the NTT implementation.