ARROW-13882: [C++] Improve min_max/hash_min_max type support

[lidavidm]

This adds support for non-nested types to both, except for MonthDayNanoInterval (which is not really sortable). hash_min_max additionally lacks binary/string-like types as they require a different approach (I will file a followup).

[github-actions]

https://issues.apache.org/jira/browse/ARROW-13882

[lidavidm]

In light of the proposal to clarify how DayTime intervals are handled some changes here should probably be reverted.

[lidavidm]

(Removed support for DayTimeInterval since it's only partially orderable.)

[pitrou]

Very nice, thank you. Just a couple questions and nits.

[pitrou]

I'm a bit surprised. This should already be checked by the ValidateArray top-level function, no?

[lidavidm]

Hmm. Odd, I do remember getting a crash for this reason here. But there is indeed a check:

arrow/cpp/src/arrow/array/validate.cc

Lines 367 to 374 in 87b2fcd

	if (type.id() != Type::EXTENSION) {
	if (data.child_data.size() != static_cast<size_t>(type.num_fields())) {
	returnStatus::Invalid("Expected ", type.num_fields(),
	" child arrays in array "
	"of type ",
	type.ToString(), ", got ", data.child_data.size());
	}
	}

[pitrou]

It seems this will generate different template specializations for types with the same underlying c_type. Is it desirable?
(note that GetValues is called with the output type independently from the ArrowType parameter)

[lidavidm]

I changed this and the one below to use PhysicalType. I also replaced GetValues with MakeScalar (since now the static type doesn't match the runtime type).

[pitrou]

Should this be else if? Presumably value cannot be bother smaller than the min and greater than the max, and it would spare a string comparison.

[pitrou]

Update message here as well?

[pitrou]

Ha :-)

[pitrou]

Same remark as for the scalar aggregations: this will generate a separate kernel per logical type?

[pitrou]

Hmm, I don't see any interval in the type definition above. Is this comment outdated?

[lidavidm]

I fixed the comments/docs.

[pitrou]

Neither is day-time, right?

[pitrou]

Same question.

[pitrou]

+1, thank you @lidavidm