Optimize swizzle_dyn AVX2 implementation

[Shnatsel]

llvm-mca shows:

• no change on Haswell and Broadwell
• 20% improvement on Skylake when using the 512-bit decomposition from Optimize swizzle_dyn for double the native vector width #540, no change otherwise
• 23% improvement on Zen 1, Zen 2 and Tiger Lake
• 7-20% improvement on Zen 3

End-to-end benchmarks on a toy base64 implementation using 512-bit shuffles improve by 7% to 13% on my Zen4 workstation, depending on how you count.

Benchmark details

$ RUSTFLAGS='--cfg disable_dispatch_avx512' cargo bench  --bench codec -- 1048576
    Finished `bench` profile [optimized] target(s) in 0.02s
     Running benches/codec.rs (target/release/deps/codec-0676c1840143d734)
encode/1048576          time:   [103.20 µs 103.31 µs 103.45 µs]
                        thrpt:  [9.4398 GiB/s 9.4524 GiB/s 9.4631 GiB/s]
                 change:
                        time:   [−11.900% −11.804% −11.685%] (p = 0.00 < 0.05)
                        thrpt:  [+13.231% +13.384% +13.507%]
                        Performance has improved.
Found 13 outliers among 100 measurements (13.00%)
  12 (12.00%) high mild
  1 (1.00%) high severe

decode/1048576          time:   [163.61 µs 163.63 µs 163.65 µs]
                        thrpt:  [5.9674 GiB/s 5.9682 GiB/s 5.9689 GiB/s]
                 change:
                        time:   [−7.6401% −7.5954% −7.5538%] (p = 0.00 < 0.05)
                        thrpt:  [+8.1710% +8.2197% +8.2721%]
                        Performance has improved.
Found 6 outliers among 100 measurements (6.00%)
  4 (4.00%) high mild
  2 (2.00%) high severe

This is a port/upstreaming of my work on fearless_simd: linebender/fearless_simd#276

Disassembly and llvm-mca measurements are from fearless_simd because this repo does not include instructions on how to disassemble and benchmark it (as opposed to the std::simd re-export of its older version).

Correctness of this formulation was verified using random testing against the scalar reference on 100k inputs: linebender/fearless_simd@4e8a56a
On top of the usual unit tests.