Implement HOST_UDF aggregation for reduction and segmented reduction

[ttnghia]

Following #17592, this enables HOST_UDF aggregation in reduction and segmented reduction, allowing to execute a host-side user-defined function (UDF) through libcudf aggregation framework.

Closes #16633.

[PointKernel]

Using a static_assert instead of SFINAE might offer clearer information in cases of misuse of this constructor, for example:

template<typename T>
data_attribute (T value_) : value {value_} {
  static_assert(std::is_same_v<T, reduction_data_attribute> ||
                std::is_same_v<T, segmented_reduction_data_attribute> ||
                std::is_same_v<T, groupby_data_attribute>,
                "Unsupported aggregation data attribute");
}

[PointKernel]

When used exclusively for groupby, the input data types are still manageable. However, they have now become somewhat challenging to control. This was also the primary difficulty I encountered when trying to understand the use of host UDF in the initial version of your design. For example, the implicit connection between the optional input scalar and the initialization value of a reduction is ambiguous to follow.

Now looking back, I think the core issue lies in using an enum to represent the data attributes required for aggregation. For example, an enum with values like GROUPED_VALUES and GROUP_OFFSETS should signify that only one of these options is valid in a given context. However, the fact that your documentation example requires both for a custom UDF violates the fundamental purpose of an enum, which is to define mutually exclusive states. Also, data_attribute is too vague, as it could refer to a (partial) groupby, a (partial) reduction, a (partial) segmented reduction, or intermediate aggregations.

[ttnghia]

How about making two tiers of base classes?

// For type erasure: storing `std::unique_ptr<host_udf_base>` by the same way in the
// internal implementation class
struct host_udf_base {};

// define data attributes, required functions etc.
struct reduction_host_udf_base: host_udf_base { };
struct segmented_reduction_host_udf_base: host_udf_base { };
struct groupby_host_udf_base: host_udf_base { };

The users will subclass from each of these 3 base classes separately. These base class will not mix their data at all.