Perf. regression of min over packed_simd

[jorgecarleitao]

There is a 25% perf regression when converting packed_simd to core_simd.

The implementations here the same mutatis mutantis, but their performances are not:

core_simd_min 2^20 f32     [286.86 us 289.22 us 292.03 us]
packed_simd_min 2^20 f32   [230.50 us 234.12 us 238.86 us]
nonsimd_min 2^20 f32       [245.75 us 249.19 us 254.00 us]
naive_min 2^20 f32         [2.8560 ms 2.8721 ms 2.8885 ms]

In particular, it seems more efficient to write the code without std::simd (nonsimd_min) than with it (core_simd_min).

with target-cpu=native:

core_simd_min 2^20 f32     [376.98 us 378.40 us 379.72 us]
packed_simd_min 2^20 f32   [181.77 us 182.95 us 185.05 us]
nonsimd_min 2^20 f32       [185.89 us 186.35 us 186.83 us]
naive_min 2^20 f32         [2.0208 ms 2.0274 ms 2.0341 ms]

which is an even larger difference

Context

We are considering migrating from packed_simd to std::simd and observed this regression on our benchmarks. jorgecarleitao/arrow2#747 for details.

[bjorn3]

This is probably #193.

[jorgecarleitao]

The conclusion on zulip is that we are not yet using intrinsics for min