Support KoalaBear field as a faster alternative to BabyBear

Motivation

Ceno currently supports BabyBear as one of its small-prime finite fields (alongside Goldilocks).
KoalaBear is a recently proposed 31-bit prime field with prime
p = 2^31 - 2^24 + 1 that offers significant performance advantages over BabyBear
(p = 2^31 - 2^27 + 1).

Why KoalaBear is faster

• Better hardware utilization: KoalaBear's prime 2^31 - 2^24 + 1 enables more efficient
  Montgomery/Barrett reduction on modern CPUs compared to BabyBear.
• Faster multiplication: Benchmarks from Plonky3, SP1, and other ZK frameworks show KoalaBear
  field multiplication is measurably faster than BabyBear on x86-64 and ARM architectures.
• Same security level: KoalaBear is also a 31-bit field with similar security and extension
  field properties (a degree-4 extension exists), so it is largely a drop-in replacement.
• Ecosystem adoption: KoalaBear has been adopted by several leading ZK projects as a direct
  replacement for BabyBear (Plonky3, SP1, etc.).

Proposal

1. Add KoalaBear support in ff_ext, providing a KoalaBearExt4 type analogous to the
   existing BabyBearExt4.
2. Expose KoalaBear as a FieldType option in the CLI (--field koalabear), alongside
   babybear and goldilocks.
3. Benchmark KoalaBear vs BabyBear within Ceno to quantify the speedup and guide users toward
   the faster option.
4. Consider making KoalaBear the default field, or optionally deprecating BabyBear while
   retaining it for backwards compatibility.

References

• KoalaBear paper: https://eprint.iacr.org/2024/278
• Plonky3 KoalaBear implementation: https://github.com/Plonky3/Plonky3/tree/main/koala-bear
• SP1 adoption of KoalaBear: https://blog.succinct.xyz/sp1-v4/

Start by reading the existing BabyBearExt4 implementation in ff_ext and the FieldType handling for the CLI's --field option. Review the KoalaBear paper and the referenced Plonky3 implementation before deciding how the extension field and compatibility choices fit Ceno. Done means KoalaBear support, CLI exposure, and comparative benchmarks are available, with the default-field decision addressed.