LLVM introduces timing-vulnerable optimizations into ML-KEM and ML-DSA operations intended to be constant-time

Summary

## Summary

On targets that do not have a conditional move/select operations, LLVM compiles the internal `conditional_sub_u32()` function, into one that will contain a conditional jump on secret-dependent data, which could leak information about encapsulation and signing keys.

`conditional_sub_u32()` is used throughout the field arithmetic of both ML-KEM and ML-DSA.

## Impact

The variable-time jumps have been observed for ML-KEM and ML-DSA on: - `i586-unknown-linux-gnu` - `s390x-unknown-linux-gnu`

Additionally, ML-DSA sees the vulnerable LLVM optimization on `x86_64-unknown-linux-gnu`, for `hazardous::dsa::mldsa44` and `hazardous::dsa::mldsa87`.

The leakage this LLVM optimization introduces could enable recovery of cryptographic secrets. The practicality of exploiting the leakage depends on the compiler, target CPU used and how "close" to the process using the vulnerable functions the attacker is.

## Resolution

Upgrade to `v0.17.16` or `v0.18.1`. `conditional_sub_u32()` has been re-written to take conditional-move/jump ops availability into account.