[FEA] Better control over the output dtype in aggregations

[wence-]

Is your feature request related to a problem? Please describe.

For the cudf-polars work, I'd like to match dtypes with polars where possible, preferably without casting the result of a libcudf call post-hoc if the interface in theory supports specifying an output type.

For whole-frame aggregations (cudf::reduce) although one is able to specify an output_dtype, this is not obeyed for a number of aggregations. Specifically:

• MEDIAN (always returns the datatype matching double)
• NUNIQUE (always returns the datatype matching cudf::size_type)
• QUANTILE (always returns the datatype matching double)

The same is true of many grouped aggregations.

Describe the solution you'd like

I'd like that aggregations could support output dtype as specified by the user.

Describe alternatives you've considered

post-hoc unary casting of the result, but this is yet another kernel launch, and produces more memory overhead.

[TomAugspurger]

#19154 highlighted a behavior difference in the streaming executor where we get the expected output for n_unique, but only via the cast we wanted to avoid in this issue. My understanding is that those casts are needed regardless.

In #19146, we'll need to skip the expected mismatch in the test at

cudf/python/cudf_polars/tests/expressions/test_agg.py

Lines 75 to 79 in 14f5cde

	# https://github.com/rapidsai/cudf/issues/15852
	check_dtypes = agg not in {"n_unique", "median"}
	if not check_dtypes and q.collect_schema()["a"] != pl.Float64:
	with pytest.raises(AssertionError):
	assert_gpu_result_equal(q)

.