Support for 32-bit word width

[yoaveshel]

Are there any plans to support 32-bit words? I rewrote the division algorithm for a 32-bit machine and got an 18% speedup (testing was done on a risc-v simulator), so it seems there are some performance to be gained here.

[DaniPopes]

Note that this would be a breaking change, but it makes sense to have machine-word-sized limbs instead of u64.

[nazar-pc]

I would actually like 32-bit limbs even on 64-bit. The use case is SPIR-V target, which is technically 64-bit, but requires special capability in practice, which is not supported by some GPUs, so I need u64 that is a [u32; 2] and u128 that is [u32; 4].