Vectorize minmax on each code path

[AlexGuteniev]

Resolves #4660

(Turns out that #4660 is completely orthogonal to #4453. Initially I thought that the "prefer values" branch is inherently vectorization friendly or can be made such, but it is not true. as that branch reduces comparisons number, whereas vectoriztion-friendly loop should be naive)

[AlexGuteniev]

Benchmark	main	Time
bm<uint8_t, 8021, Op::Both_val>	4126 ns	100 ns
bm<uint16_t, 8021, Op::Both_val>	4200 ns	229 ns
bm<uint32_t, 8021, Op::Both_val>	361 ns	362 ns
bm<uint64_t, 8021, Op::Both_val>	2938 ns	2889 ns
bm<int8_t, 8021, Op::Both_val>	3715 ns	102 ns
bm<int16_t, 8021, Op::Both_val>	3156 ns	235 ns
bm<int32_t, 8021, Op::Both_val>	363 ns	364 ns
bm<int64_t, 8021, Op::Both_val>	3120 ns	3123 ns
bm<float, 8021, Op::Both_val>	1329 ns	1318 ns
bm<double, 8021, Op::Both_val>	2677 ns	2560 ns

[StephanTLavavej]

I'm mirroring this to the MSVC-internal repo - please notify me if any further changes are pushed.

[StephanTLavavej]

Thanks for closely monitoring performance and making sure all cases are covered! 🚀 😻 💯