GH-49003: [C++] Don't consider out_of_range an error in float parsing

[Alvaro-Kothe]

Rationale for this change

This PR restores the behavior previous to version 23 for floating-point parsing on overflow and subnormal.

fast_float didn't assign an error code on overflow in version 3.10.1 and assigned ±Inf on overflow and 0.0 on subnormal. With the update to version 8.1, it started to assign std::errc::result_out_of_range in such cases.

What changes are included in this PR?

Ignores std::errc::result_out_of_range and produce ±Inf / 0.0 as appropriate instead of failing the conversion.

Are these changes tested?

Yes. Created tests for overflow with positive and negative signed mantissa, and also created tests for subnormal, all of them for binary{16,32,64}.

Are there any user-facing changes?

It's a user facing change. The CSV reader on version libarrow==23 was assigning them as strings, while before it was parsing it as 0 or +- inf.

With this patch, the CSV reader in PyArrow outputs:

>>> import pyarrow
>>> import pyarrow.csv
>>> import io
>>> table = pyarrow.csv.read_csv(io.BytesIO(f"data\n10E-617\n10E617\n-10E617".encode()))
>>> print(table)
pyarrow.Table
data: double
----
data: [[0,inf,-inf]]

Closes #49003

• GitHub Issue: [Python] CSV reader returns different values in 23.0.0 #49003

[github-actions]

⚠️ GitHub issue #49003 has been automatically assigned in GitHub to PR creator.

[WillAyd]

lgtm @pitrou

[WillAyd]

Nit on readability - I think it would help to break up this line. Something like:

const auto is_allowed_ec = res.ec == std::errc() || res.ec == std::errc::result_out_of_range;
return is_allowed_ec && res.ptr = s + length;

[Alvaro-Kothe]

Definitely it's better. I also took the liberty to store the pointer comparison in an auxiliary variable in favor of readability.