<numeric>: Make gcd() faster with constexpr bitscan intrinsics

[StephanTLavavej]

Our gcd() implementation contains the following workaround:

STL/stl/inc/numeric

Lines 847 to 860 in 58bb49d

	template <class _Unsigned>
	constexpr unsigned long _Stl_bitscan_forward(_Unsigned _Mask) noexcept {
	// find the index of the least significant set bit (_BitScanForward isn't constexpr... yet :))
	static_assert(is_unsigned_v<_Unsigned>, "Bitscan only works on bits");
	unsigned long _Count = 0;
	if (_Mask != 0) {
	while ((_Mask & 1U) == 0) {
	_Mask >>= 1;
	++_Count;
	}
	}
	return _Count;
	}

For #25 (WG21-P0553 <bit> Rotating And Counting Functions), we're getting compiler support for constexpr bitscans (with different intrinsic names). We should use this in gcd().

Note that because <bit> is a C++20 header, we can't include it from <numeric>, nor can we use its non-_Ugly functions to implement C++17 features like gcd(). However, we can use the compiler intrinsics directly, or define _Ugly wrappers for them in centralized headers.

Also tracked by Microsoft-internal VSO-597250.