batch_key of highly_variable_genes not working

Hey!

I'm currently having an issue with the batch_key functionality of sc.pp.highly_variable_genes introduced in https://github.com/theislab/scanpy/pull/622 by @gokceneraslan.

If I try setting batch_key, I get a TypeError:

---------------------------------------------------------------------------
TypeError                                 Traceback (most recent call last)
 in 
----> 1 sc.pp.highly_variable_genes(adata, batch_key="sample", flavor='cell_ranger', n_top_genes=4000, inplace=False)

~/.conda/envs/sc-tutorial/lib/python3.7/site-packages/scanpy/preprocessing/_highly_variable_genes.py in highly_variable_genes(adata, min_disp, max_disp, min_mean, max_mean, n_top_genes, n_bins, flavor, subset, inplace, batch_key)
    336             dtypes.append([('highly_variable_nbatches', int),
    337                            ('highly_variable_intersection', np.bool_)])
--> 338         return np.rec.fromarrays(arrays, dtype=dtypes)

~/.conda/envs/sc-tutorial/lib/python3.7/site-packages/numpy/core/records.py in fromarrays(arrayList, dtype, shape, formats, names, titles, aligned, byteorder)
    606 
    607     if dtype is not None:
--> 608         descr = sb.dtype(dtype)
    609         _names = descr.names
    610     else:

TypeError: data type not understood

I have tried with multiple categorical columns and it worked with none of them.

Start in scanpy/preprocessing/_highly_variable_genes.py at the highly_variable_genes path handling batch_key and the np.rec.fromarrays call shown in the traceback. Reproduce the failure with a categorical batch_key and verify that the function completes without the TypeError and returns the expected highly-variable-gene results.