Faster bn254 syscalls

[cryptopapi997]

Inspired by #408 , we can steal adapt these ideas to optimize the bn254 syscalls as well, since those all overwrite the full memory buffer as well. For the compression syscalls that return a fixed size array, we can utilize the same MaybeUninit trick. For the addition, multiplication, and pairing syscalls we can alloc a vec once at the beginning instead of allocing an array and then converting that to a vec. This results in the following CU diffs:

operation	Δ CU
add	-27
mul	-27
pairing	-13
g1_compress	-4
g2_compress	-8
g1_decompress	-8
g2_decompress	-22

Note: I didn't apply this to the deprecated bn254 syscalls, but can do as well.
Note 2: Technically these optimizations could be applied inside the actual syscall implementation too instead of the wrapper functions interfacing with them for a couple extra CUs. However, this seemed more low risk, and idk if a SIMD would apply in that case so I went with this for now.

[joncinque]

Thanks for your contribution, looks great to me! Just the one comment, @samkim-crypto can you also take a look?

[joncinque]

Noting for anyone watching, that the Vec::with_capacity -> as_mut_ptr() -> set_len() flow is very well demonstrated here: https://doc.rust-lang.org/std/vec/struct.Vec.html#method.as_mut_ptr

[samkim-crypto]

Just one small clarification below, but everything looks fine to me!