convert vello_cpu to use #[simd]

[Shnatsel]

demonstrates how linebender/fearless_simd#347 applies to vello_cpu

x86 build times are unchanged, surprisingly.

No performance regressions from this PR. The initial conversion introduced some, but sticking #[inline(always)] on 3 functions fixed it - that's commit de2e4c1

I'm seeing performance improvements for fills:

fine/fill/opaque_short_u8_sse4_2
                        time:   [5.9306 ns 5.9947 ns 6.0566 ns]
                        change: [-19.952% -19.305% -18.819%] (p = 0.00 < 0.05)
                        Performance has improved.

fine/fill/opaque_short_u8_avx2
                        time:   [4.4478 ns 4.4540 ns 4.4601 ns]
                        change: [-32.146% -30.706% -29.160%] (p = 0.00 < 0.05)
                        Performance has improved.

fine/fill/opaque_long_u8_avx2
                        time:   [14.128 ns 14.207 ns 14.296 ns]
                        change: [-41.763% -41.529% -41.232%] (p = 0.00 < 0.05)
                        Performance has improved.

fine/strip/solid_long_u8_avx2
                        time:   [67.304 ns 67.348 ns 67.394 ns]
                        change: [-20.794% -20.736% -20.670%] (p = 0.00 < 0.05)
                        Performance has improved.

[LaurenzV]

Seems like it somehow negatively affects WASM size for vello_cpu? Any idea why?

[Shnatsel]

Hmm, 20% raw size increase but 3.8% increase gzipped; probably added something repetitive that gzip can easily remove.

I'll see what I can do to investigate.

[Shnatsel]

Codex bisected the regression to linebender/fearless_simd@ff8ee04

I've reverted that commit, let me re-run the benchmarks

[Shnatsel]

fine/fill/opaque_long_u8_avx2
                        time:   [13.623 ns 13.624 ns 13.625 ns]
                        change: [-43.012% -42.974% -42.939%] (p = 0.00 < 0.05)
                        Performance has improved.

This performance improvement stays, the other fill improvements are gone. I'll see what I can do about that. Also, one more regression appeared that I'll need to address.

[Shnatsel]

Fixed pixmap/unpremultiply/opaque benchmark regression.

The short AVX2 paths vary in timings between runs, so the earlier 20% improvement was spurious.

Results as of 1d34528: fine/fill/opaque_long_u8_avx2 improves by 40%, all other run-by-default benchmarks are neutral.