ML-DSA Parameter Set Monomorphization with Macros

[jschneider-bensch]

The motivation for this PR is to mostly get rid of allocations below the level of ml_dsa_generic. To this end, const generics are replaced by a macro-based manual monomorphization to the different parameter sets. In doing this, I've attempted to simplify parameter set specific constants, as well as platform instantiation and multiplexing.

Essentially ml_dsa_generic, constants, the instantiations in ml_dsa_generic::instantiation::{portable, neon, avx2} and ml_dsa_generic::multiplexing now have (macro-generated) submodules for each parameter set which should be guarded by their respective feature flag. They contain:

• constants::ml_dsa_44/65/87: Parameter set specific constants go here. The constants module also defines const fns for deriving everything but the basic constants in the parameter set specific submodules.
• ml_dsa_generic::ml_dsa_44/65/87: A version of the ml_dsa_generic::generic module where the parameter specific constants from constants are pulled in and derived constants are derived using the const fns from constants. The module also defines type aliases MLDSA{44/65/87}SigningKey, MLDSA{44/65/87}VerificationKey, etc. This is achieved by the ml_dsa_parameter_sets proc macro.
• instantiations::{portable, neon, avx2}::ml_dsa_44/65/87: The platform specific instantiations of the versions from ml_dsa_generic. In the case of avx2that includes the application of thetarget_platform` cfg.
• multiplexing::ml_dsa_44/65/87`: Runtime multiplexing, reverting to a portable version if platform-specific version was not compiled in.

I've verified that F* extraction still works and laxes. C extraction pending further updates.

[karthikbhargavan]

This is a big one! Left a few comments, but happy for most of these being follow-ups.

However, let's document the general style guidelines implemented here, so that new code can also follow this pattern. I am adding a comment to the PR noting this.

[karthikbhargavan]

This PR does a bit more cleanup than just the const generic monomorphization.
Let's try to document some of these style decisions so that we can be uniform in the codebase (and backport to ML-KEM and SHA-3 when we touch that code again.). Here are some things I noticed:

• All struct and array arguments to functions are now borrows
• We uniformly use &mut arguments instead of returning a new array or struct
• We uniformly use slices (with implicit lengths) instead of fixed-size arrays
• We eliminate const generic arguments and replace them by real arguments (when truly needed)
• In several places, we replace calls to complex iterators like enumerate with simpler for loops over ranges.

[jschneider-bensch]

You're right, there's more here than the monomorphization. Thanks for noting down the style changes!

[franziskuskiefer]

I think we should simplify the plumbing for the variants a little to make it easier to add nice APIs. But that's for follow-ups. Let's get this in.