Use MaybeUninit in place of 0-initialized u8 arrays

[apoelstra]

Originally posted by @Kixunil (in a particular case) in #123 (comment)

There may be other cases -- typically with code Tobin or I touched we were not looking for unsafe optimizations.

[Kixunil]

I often opted for ArrayVec to avoid unsafe but if it's shown in benchmarks to cause problems then I'm fine with MaybeUninit. Still, it's really weird that it happens at all.

[apoelstra]

I suspect the perf hit comes from ArrayVec being generic being generic on T. I think if there were a u8-specific one (or maybe just T: Copy) (or even T: !Drop?) it would be faster.

It may also be that ArrayVec does length accounting which we don't need -- we always have fixed sizes that we overwrite completely (and, I think, have no unwinding branches between creating the object and overwriting it).

I dunno. Everything I just said, I'd have expected the compiler to take care of when using ArrayVec<u8>.

[apoelstra]

Oh -- another thing we should consider is using core::array::from_fn! This is safe, allowed by MSRV since 1.63, and should encapsulate exactly the logic we want.

[Kixunil]

Well, I'd expect post-mono/inlining optimizations to see that the lengths are the same and delete those branches/increments. I don't believe it has anything to do with it being generic.