Cudf: [FEA] Port contents of functions.h to new cudf::column and associated types

[harrism]

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

Now that #2207 is merged we need to start porting functionality to use it. This issue covers functionality defined in cudf/functions.h.

Describe the solution you'd like

Specifically:

• [x] gdf_order_by (already implemented in #2207, so this is simply a header/namespace change)
• [x] gdf_context_view (not sure what to do with this -- need to look at usage
  Also, functionality in functions.h largely uses the legacy "gdf" file structure and
• [x] gdf_valid_allocation_size and gdf_num_bitmask_elements won't be needed once the transition is complete, so for now they should just move to a header in the legacy folder and legacy namespace.
• [x] nvtx functions: useful to keep, should just be updated to throw exceptions and moved to their own header and in the cudf::utility namespace or similar.
• [x] Error handling functions gdf_error_get_name etc.
• [x] gdf_extract_datetime_second and similar function may need to be handled in a separate refactor of date time functions.
• [x] gdf_digitize
• [x] gdf_transpose
• [x] gdf_hash
• [x] gdf_hash_partition
• [x] gdf_to_dlpack
• [x] gdf_from_dlpack

Not all of the above needs to be done in a single PR. In fact, I recommend grouping it logically into multiple PRs to make reviewing and merging easy (and to enable collaboration).

[jrhemstad]

gdf_order_by can be removed entirely.

gdf_error_get_name and error handling functions as well as gdf_context_view can/should just be removed all together.

gdf_valid_allocation_size and gdf_num_bitmask_elements are superseded by https://github.com/rapidsai/cudf/blob/branch-0.10/cpp/include/cudf/null_mask.hpp#L47

[trevorsm7]

I started on this today

[harrism]

Please follow the two-step approach (Start with a PR that only moves files to legacy and points include paths at them)

[trevorsm7]

Why are the gdf_nvtx_range_* functions declared in functions.h instead of nvtx_utils.h? These don't have anything to do with columns, but I could update them to CUDF_EXPECTS instead of returning gdf_error.
EDIT: Oh, I missed this at the bottom of your list, Mark. Will do just that.

[trevorsm7]

First step here: https://github.com/rapidsai/cudf/pull/3070

[jrhemstad]

Why are the gdf_nvtx_range_* functions declared in functions.h instead of nvtx_utils.h?

Because _all_ external APIs used to be defined in functions.h :)

[trevorsm7]

datetime_ops functions are covered by https://github.com/rapidsai/cudf/pull/3201

[trevorsm7]

digitize is being deprecated by upper_bound and lower_bound: https://github.com/rapidsai/cudf/issues/3258

[trevorsm7]

All PRs merged. Closing!

[harrism]

Congrats and great work @trevorsm7, this was a tall order.