Add concat_swizzle_dyn and concat_swizzle_dyn_precise

[Shnatsel]

1. Add concat_swizzle_dyn and concat_swizzle_dyn_precise
2. Change swizzle_dyn_precise SSE4.2 128-bit formulation; it's one fewer instruction, slightly worse latency, but much better optimization when larger vectors are decomposed into it.
3. Use select_unpredictable stabilized in 1.88 in scalar fallbacks of concat_swizzle_dyn_precise and swizzle_dyn_precise.

Rationale for the branchless formulation: cmov is far from a silver bullet and is often a bad idea. In my benchmarks this is 20% to 40% slower than branches for swizzle_dyn_precise if the indices are all out of range, but avoids a 5x (400%) performance degradation on unpredictable indices. If you know your indices are predictable, you don't use _precise in the first place, so tuning for unpredictability seems fine. Most importantly, removing branches helps autovectorizer a lot, so we get more reliable autovectorization on e.g. RISC-V, which is the primary use case.

Implementation-wise, NEON and AVX-512 are straightforward with native-ish instructions, SSE4.2 uses the same decomposition idea as swizzle_dyn_precise, and AVX2 needs babysitting as always.

The first commit is squashed because I did a bunch of experiments with different formulations, e.g. a specialized one for 256-bit AVX2 that reduces shuffle port pressure on haswell and doesn't regress recent AVX2 much if at all, but in the end decided those are not worth the extra complexity.

I'm not thrilled by the lhs.concat_swizzle_dyn(rhs, indices) API but I don't see how to do better without it getting really awkward in other ways: concat_swizzle_dyn(simd, lhs, rhs, indices) isn't any better and breaks convention, and indices.concat_swizzle_dyn(lhs, rhs) runs into issues with matching vector lengths.

[Shnatsel]

@Novum please let me know if this looks good to you

[Novum]

Looks good for me for those operations