Use iodata in the internal state of StringIO

[sabiwara]

An attempt at the 3rd point mentioned here.

I haven't done proper benchmarks yet, but it seems actually slower by ~1s when trying this with 32 - will look into it later this week.

[josevalim]

Btw, to help us find bugs, I would default the initial state to an empty list.

[sabiwara]

I've been struggling with this one, the concatenation always seems faster when running the benchmark mentioned above (~5.6s on main, ~7.0s on this branch). My assumption is that this could be due to this optimization in OTP 26.

I updated to use a reverse list rather than a [acc | binary], because this seems faster in individual concatenation benchmarks (at the cost of more memory though).

But even there it isn't an overwhelming improvement over string concatenation which is still occasionally faster, and I can't observe any improvement in StringIO itself.

Maybe I'm missing something, but I'm tempted to close this one and consider StringIO as already close to optimal 😅

[josevalim]

Thank you for exploring! I am ok with closing this, your call!

[sabiwara]

Thanks @josevalim 💜
Will close, we can always reopen if we find new ideas or other benchmarks where this approach works better.