More minmax vectorization

[StephanTLavavej]

Followup issue after merging #2447.

• Vectorize ranges::min_element, ranges::max_element, ranges::minmax_element
  • This should be similar to vectorizing ranges::count, as we should be able to call the same vector_algorithms.cpp functions after dealing with the iterator/sentinel. (Unlike ranges::find, we don't need to worry about unbounded ranges.)

And as @AlexGuteniev explained, we can go even further:

• Vectorize ranges::min, ranges::max, ranges::minmax
  • These algorithms return values instead of iterators, so they can be simpler and faster, and "will consist only of vertical max and one reduction".
• Use AVX2.

#2439 tracks extending the optimization to floating-point types.

[AlexGuteniev]

An observation.
We have an implementation that uses minmax_element and implementation that just goes just iteratively, saving minmax values. The later might auto-vectorize, in this case we can avoid separately compiled impl

STL/stl/inc/xutility

Lines 7002 to 7011 in ed8150e

	if constexpr (forward_range<_Rng> && _Prefer_iterator_copies<iterator_t<_Rng>>) {
	return static_cast<range_value_t<_Rng>>(*_RANGES _Min_element_unchecked(
	_STD move(_UFirst), _STD move(_ULast), _Pass_fn(_Pred), _Pass_fn(_Proj)));
	} else {
	range_value_t<_Rng> _Found(*_UFirst);
	while (++_UFirst != _ULast) {
	if (_STD invoke(_Pred, _STD invoke(_Proj, *_UFirst), _STD invoke(_Proj, _Found))) {
	_Found = *_UFirst;
	}
	}