Optional fillna in exclude_all_nan_cols.

The default behaviour has been changed to be of more general use.
The previous default was to fill the NaNs not belonging to
all NaN columns with 0, which was suitable for standardized usage
biases (which are expected to be centered on 0), but we might prefer
to keep NaNs if we want to later replace those values with something
else when dealing with other usage metrics than biases. If a single
filling value is desired, it can be set using argument fill_other_nas.

Implementation detail: The fillna still happens, but with a default
value of np.nan, which hopefully should be OK.