Switch to std::simd

[RazrFalcon]

TODO:

• [ ] f32 pipeline is a bit broken. Specifically gradients and patterns. No idea why. Maybe due to rounding.
• [ ] f32 pipeline is up to 40% slower on aarch64. Should start with fixing color_burn. Maybe Mask::select is not getting vectorized?
• [ ] Bench on x86.
• [ ] Make sure output on x86 and aarch64 is identical.
• [ ] Figure out what to do with simd::StdFloat and no_std.
• [ ] Figure out what to do with the old f32x8::recip_fast method.
• [ ] Wait until std::simd will become stable...

Some methods in the u16 pipeline are up to 20% faster on aarch64, which is very good.

Gradients in the u16 pipeline are slightly different because of explicit mul_add instead of a manual f * m + a.

Closes #59

[notgull]

Not sure if you're planning on releasing this soon, but I would be against making tiny-skia a nightly-only package. It's currently relied on by winit and theo, which are not nightly-only and would likely be opposed to becoming nighly-only.

[RazrFalcon]

@notgull

Wait until std::simd will become stable...

[notgull]

Ah, my bad, I missed that option. Thanks!