[FEA] Replace use of cudf.datasets/cudf.testing utils for generating cuDF classic test data with standard constructors

Some cuDF classic unit tests construct input data using methods from

* `cudf.datasets`
* `cudf.testing._utils` e.g. `gen_rand`
* `cudf.testing.dataset_generator`

and while these methods allow creating input test data by specificity certain data characteristics, IMO

1. They are more opaque than using public cuDF Python APIs
2. Testing "data charactistics" for correctness should ideally happen in libcudf, the cuDF classic test suite should focus on testing compatibility with pandas (which understandably sometimes involves testing data characteristics)

The task is to replace usages of the above with

1. Inline the creation of input data with a public cuDF API e.g. `data = cudf.DataFrame(...)`
2. A pytest fixture that creates "generic" data if applicable