Resolves #6020

⚙️ Optimization

For usual types, compare for equality, instead of equivalence:

• For integers this allows using vectorization path for <flat_set>
• For strings this allows way cheaper comparison, as suggested by @StephanTLavavej
• For anything if ranges::less or ranges::greater are used, as @cpplearner suggested, other types with string-like situation may benefit too

@frederick-vs-ja suggested to take advantage of sorted range, and always perform only one comparison. But we need other unique algorithm for that.

⏱️ Benchmark results

🥈 Results interpretation

Most of the time is spent on sorting the inserted data, yet some is spent on making it unique, so the improvement can be observed, but it is limited.

For strings this played out well.

For integers there's no improvement seen with the current benchmark. flat_map of integers is not improved, because the compiler is able to replace equivalence with equality on its own; and even if it wouldn't, the difference would be very minor anyway.

flat_set of integers goes the vectorized path, but the benchmark needs more pieces to make vectorization beneficial. And with more pieces, the sorting would consume even greater portion of time, due to above-linear complexity.

So looks like this is useless for integers, but strings still benefit from this.