Support categorical variables with CSVs

[esafak]

It would be nice to be able to read CSVs with categorical variables using read_csv's dtype parameter instead of casting the columns after the fact.

[TomAugspurger]

I'm not opposed to this in principle, but I think the API will necessarily be clunky. Would we require (or allow) the user to specify all categories in the call to read_csv.

@esafak we do support categoricals in read/write_hdf if that's an option for you (it may not be).

[esafak]

Can't we already declare the dtypes of selected columns? I thought the problem was limited to categoricals, but if not, please expand my request to all dtypes.

[TomAugspurger]

You can specify the types. I was just thinking

pd.read_csv('file.csv', dtypes={'A': np.int64, 'B': pd.CategoricalDtype(['cat1', 'cat2', 'cat3'])})

which means you'd need to know all the categories up front. Or we infer them and you'll need to check that they're aren't any surprising categories.

[sinhrks]

Nice workaround, but I think it is still nice to support category arg.

As a first step, how about converting the specified columns to Categorical after parsing? Though it is very nice to have optimized IO logic...