<bitset>: Investigate further performance improvements

[StephanTLavavej]

Followup to #3838 which is adding a benchmark - thanks @achabense!

• @AlexGuteniev suggested using basic_string::resize_and_overwrite, creating an _Ugly version for unconditional use internally. (We generally avoid adding totally novel secret machinery to Standard classes as it can become a maintenance burden, but simply having _Ugly names to access Future Technology is easy and common.)
• @AlexGuteniev also suggested investigating whether branchless codegen for assigning _Elem0 vs. _Elem1 would be superior. (Note: we cannot assume that their values are consecutive, they could be 'M' and 'E'.)
• I suggest investigating SIMD, as this would seem to be highly amenable to vectorization.

Regardless of what's investigated, I recommend profiling each change in isolation.

[AlexGuteniev]

Vectorization approach suggested by @Alcaro :

void bits_to_str16_walrus(char* dest, uint16_t src)
{
	__m128i bits = _mm_cvtsi32_si128(src);
	bits = _mm_unpacklo_epi8(bits, bits);
	bits = _mm_unpacklo_epi8(bits, bits);
	bits = _mm_unpacklo_epi8(bits, bits);
	bits = _mm_and_si128(bits, _mm_set1_epi64x(0x8040201008040201));
	bits = _mm_cmpeq_epi8(bits, _mm_setzero_si128());
	bits = _mm_and_si128(bits, _mm_set1_epi8(0x01));
	bits = _mm_xor_si128(bits, _mm_set1_epi8(0x31));
	_mm_storeu_si128((__m128i*)dest, bits);
}

https://godbolt.org/z/hK1YaY8fK
(saving it here for now, so that it is not lost while reviewing the current PR)