ARROW-11247: [C++] Infer date32 columns in CSV

[nealrichardson]

No description provided.

[github-actions]

https://issues.apache.org/jira/browse/ARROW-11247

[nealrichardson]

My editor runs clang-format and IDK why it moved this

[nealrichardson]

Some python tests are failing with this change; I think the tests should be updated from where it assumes dates will be parsed (suboptimally IMO) as timestamps, but @jorisvandenbossche maybe you can review and weigh in.

[jorisvandenbossche]

I added a commit that updates the tests for the new behaviour, in case we decide we are OK with that.

Generally I think we should do the best inference from Arrow's point of view, and which is thus date type for a date string.

The only reason I am thinking to not do it is that, for people converting the data to pandas afterwards, dates are not that well supported (at this point) in pandas. Now, there is a to_pandas(..., date_as_object=False) keyword a user can specify to still get a datetime64 dtype in pandas instead of datetime.date objects.

[pitrou]

Please update the docs for https://arrow.apache.org/docs/cpp/csv.html#data-types and https://arrow.apache.org/docs/python/csv.html

[pitrou]

Travis-CI build: https://travis-ci.com/github/pitrou/arrow/builds/213173530