Faster to_bignum_bigint

[mitschabaude]

o1js: o1-labs/o1js#1338

This reimplements conversion of Rust field elements to Zarith bigints (Bigint.to_bignum_bigint). The conversion is used for the Hashable implementation of field elements, which previously was one of the main bottlenecks during witness generation, because plonk_constraint_system.ml uses a hashtable with field element keys.

The old conversion tested each individual bit of the Rust field. The new version tests 32 bits at once, and introduces a new bindings function test_uint32 for this purpose.

Empirically, this PR makes brings the witness generation time in a JS benchmark (Keccak) down by about 60%, from 1.7 to 0.7 seconds. The cost of hash calls, which used to be >50%, becomes negligible.

Note: IMO, this isn't the cleanest way to achieve the goal. It keeps the dependency on Zarith -- a separate, arbitrary-length bigint implementation, which is slow at least in the JS version -- just for support of a few methods like Hashable.hash. Ideally, we would implement hashing in Rust and JS and get rid of Zarith. However, that's a more impactful change and would have required more research. This here seems to be the minimal change that removes the bottleneck.

TODO: So far, only the JS backend implements test_uint32. Rust and Rust/wasm versions have to be added.

[mitschabaude]

It turns out that the same function was optimized in rampup already: #14599
I tested and confirmed that it gives the same benefit as this PR. So for the sake of simplicity I'll close this.