use a dynamic bytes swizzle for swizzle_to_front

[Dr-Emann]

Fixes #207

As the comments say, unfortunately, this leads to pretty suboptimal code unless the std library is compiled with the required target feature, and the default x86_64 target (except for macos) is not compiled with the required target feature.

See godbolt showing assembly for: default x86_64 (bad assembly), aarch64 (good assembly), and x86_64 macos (good assembly, which should also be the assembly for -Zbuild-std)

We could use a little unsafe and do something like:

#[cfg(target_feature = "ssse3")]
{
    use core::arch::x86_64::{__m128i, _mm_shuffle_epi8};
    let val_m128 = __m128i::from(val_convert);
    let swizzle_m128 = __m128i::from(swizzle_idxs);
    let swizzled_m128 = unsafe { _mm_shuffle_epi8(val_m128, swizzle_m128) };
    return u16x8::from(swizzled_m128);
}

to allow optimal code generation with just a new enough target in the caller without requiring recompiling std.

[Kerollmops]

Hey @Dr-Emann 👋

Thank you for your work. Very good explanation of the problem, as always.

I am wondering if the unsafe code you are showing is not the way to go? If the code is well documented we should go this way even if it's unsafe. What do you think? 🤔

[Dr-Emann]

Sure. I added a version which also uses runtime detection if the std feature is enabled as well, so even if the user doesn't target a cpu with ssse3 directly, they still won't have to go through the non-simd version unless they run on an ancient processor.

[Kerollmops]

Amazing work! Thank you for this 👍