BUG: CategoricalIndex.where nulling out non-categories

[jbrockmendel]

• closes #xxxx
• tests added / passed
• passes black pandas
• passes git diff upstream/master -u -- "*.py" | flake8 --diff
• whatsnew entry

ATM CategoricalIndex.where has some idiosyncratic behavior:

ci = CategoricalIndex(["a", "b", "c", "d"])
mask = np.array([True, False, True, False])

>>> ci.where(mask, 2)
CategoricalIndex(['a', nan, 'c', nan], categories=['a', 'b', 'c', 'd'], ordered=False, dtype='category')

This makes that call raise instead.

Index.where is only used in one place in reshape.merge. Might be worth deprecating+privatizing.

[jreback]

this is an odd message for this operation.

[jbrockmendel]

yah thats a side-effect of de-duplicating all the validators. the datetimelike one is nicer, so im planning on doing a pass to make the categorical one behave more like the datetimelike one. that'll be a big diff since it will involve changing a lot of tests, so planning to do that in a dedicated PR

[jreback]

also pls add a whatsnew note

[jbrockmendel]

whatsnew added + green