Use Żmij algorithm for speed up float `repr()` / `str()`

[chirizxc]

🐉 Żmij

• 🐉 Żmij Algorithm

• 🐉 Żmij benchmark

• 🐉 Żmij licence (MIT)

• dtoa-benchmark where this algorithm is benched

As I understand it, the MIT license allows us to vendor this code into CPython, so there shouldn't be any issues with that.

On my machine, this results in a 4.09× speedup for repr(x) and a 4.33× speedup for str(x)

Linked PRs

• gh-155300: speed up float repr() / str() with new Żmij algorithm #155298

[eendebakpt]

@chirizxc There is discussion on this topic at https://discuss.python.org/t/faster-float-to-string-conversions/107397 (and references there).

Zmij is indeed faster, but it is unclear there are real-world cases where it matters and just adding Zmij to cpython will increase the maintenance work