Implement the reinterpret operations without bytemuck

[valadaptive]

Stacked on top of #116, because it touches some of the codegen stuff I cleaned up in that PR. it's unfortunate that GitHub doesn't have stacked PRs.

We have the Bytes trait, which lets us cast SIMD types to and from raw bytes (currently using mem::transmute). We can use its bitcast method instead of pulling in bytemuck for the "reinterpret" operations on Fallback.

On the x86 side, we can use the _mm_cast[...] intrinsics. All the x86 integer types are __m128i or __m256i, so conversions between integer widths are no-ops.

While working on this, I noticed that there are "reinterpret signed as unsigned" ops, but no corresponding "reinterpret unsigned as signed" ops. Are the reinterpret ops worth it at this point if we have the Bytes trait?

[DJMcNab]

Initial reaction from a skim is that this seems reasonable and much more ergonomic. In #simd > ✔ Using `bytemuck`, we discussed using bytemuck for the operations which require unsafe (i.e. implementing Bytes). That would be especially neat in a post-#108 world, where everything would then be validated.

The changes in this PR are however pretty orthogonal to that. Using the appropriate SIMD intrinsic does make sense (especially if we get something like #simd > Function call ABI).
Otherwise I can't approve because of the stacking.

[valadaptive]

I think we can reintroduce bytemuck in the future to back the Bytes trait if we want to avoid transmutes (although since the types are numeric, from_ne_bytes()/to_ne_bytes() should work).

I'll request a re-review once everything else is merged.

[valadaptive]

This is ready to review now.

[LaurenzV]

Lovely, thanks!