test+fix: Add asm testing and fix core_hint_black_box

[Daniel-Aaron-Bloom]

Per this zulip discussion and rust-lang/rust#140341 and in contrast to the discussion on #107, read_volatile (and all other equivalent functions in the standard library) should provide strict fewer optimization barriers than black_box.

In service of testing this, I've added a script which makes sure the assembly instructions of a simple case are approximately the expected length across most common architectures.

Also read_volatile is never going to be const, and it would be nice to constify this library (which will hopefully be my next PR if this one is accepted).

[elichai]

Semi Related: #135 (Also touches the cfg)

[cmlsharp]

Const-ness and better "guarantees" (heuristic guarantees anyway) are the main reasons to prefer core::hint::black_box but it's worth mentioning there are minorly positive performance benefits to using core::hint::black_box as well. (though I sort of doubt this is an actual performance bottleneck for anyone).

I've been writing my own version of subtle on the side, and from my benchmarks, you get about a 1.3-1.5x speedup switching from the read_volatile based black box to the hint::black_box one simply because the read_volatile version requires calling an #[inline(never)] function which results in a bunch of unnecessary stack manipulations. That said, this takes ~50k comparisons to rise to the level of milliseconds on my laptop

The difference in ASM is something like the following:

read_volatile_select_if_eq:
        push rbp
        push rbx
        push rax
        mov ebx, esi
        mov ebp, edi
        xor edi, edi
        cmp edx, esi
        sete dil
        call subtle::black_box
        movzx eax, al
        neg eax
        xor ebp, ebx
        and ebp, eax
        xor ebp, ebx
        mov eax, ebp
        add rsp, 8
        pop rbx
        pop rbp
        ret
        
 subtle::black_box:
        mov byte ptr [rsp - 1], dil
        movzx eax, byte ptr [rsp - 1]
        ret

vs

hint_bb_select_if_eq:
        cmp esi, edx
        sete byte ptr [rsp - 1]
        lea rax, [rsp - 1]
        movzx eax, byte ptr [rsp - 1]
        neg eax
        xor esi, edi
        and eax, esi
        xor eax, edi
        ret

edit: of course the core::hint::black_box version still has some duplicate loads/stores but that's just kinda the price you pay for the optimization barrier approach I think.