Use a column to store categories, rather than a mapping

[shwina]

This PR implements one of the changes mentioned in #41, i.e.,

Add get_children() method, and store the mapping that is now in Column.describe_categorical in a child column instead. Note that child columns are also needed for variable-length strings.

Replacing the previous mapping with a child column now implies that the data buffer stores integer indices into the child column.

[vnlitvinov]

@shwina @rgommers should we maybe not try to disguise the (explicit) categories as generic "children"?

I think just replacing the type of df.describe_categorical['mapping'] from dict to Column instead of introducing of a new method could be better - it's still a mapping after all 🙃

[rgommers]

We briefly touched on this PR today, we'd like to move it forward since all interchange protocol implementers seem to be in favor of this. @vnlitvinov plans to revisit his proposal from above and add that here or open an alternative PR. Then we can try to finalize this, and then update implementations in the various libraries for it.

[shwina]

still a mapping after all

I'm not sure I quite follow; the Column is explicitly not a mapping, rather it is concretely the categories. Perhaps I'm missing something?

[vnlitvinov]

@shwina I've missed your comment, so I went ahead with the original plan we discussed last meeting and I had added a commit doing so. If you don't like the idea feel free to remove my commit.

As for it not being a mapping... it somewhat is :) it maps an integer index to a category value.

[rgommers]

After discussion on Thursday, @shwina and others were happy with the changes @vnlitvinov pushed. With one change to make: rename mapping to categories, since it's no longer a real mapping (only implicitly, the keys are 0 ... n_categories) and it would help with transitioning too.

[rgommers]

With one change to make: rename mapping to categories,

This change was made in the spec part, but not the rest of this PR. I'm inclined to push a change to pandas_implementation.py and then merge this.

[rgommers]

With one change to make: rename mapping to categories,

This change was made in the spec part, but not the rest of this PR. I'm inclined to push a change to pandas_implementation.py and then merge this.

I tried this - since the tests no longer pass with this PR, and we now have an actual Pandas implementation, it may be better to completely remove this early prototype in pandas_implementation.py.

[rgommers]

I tried this - since the tests no longer pass with this PR, and we now have an actual Pandas implementation, it may be better to completely remove this early prototype in pandas_implementation.py.

Discussed in today's call: everyone is happy with deleting this prototype code.

[rgommers]

Everyone is happy and this is now a pretty small/straightforward diff. So merging. I'll follow up with all the implementers to ensure we will actually propagate these changes to all implementations.

Thanks @shwina, @vnlitvinov, @kkraus14