ENH: nansum for all nan arrays

[ustcfdm]

Proposed new feature or change:

For the function of nansum, when all elements are nan, it returns 0. However, in some cases I want it return 'nan'. Is it possible to add an option to let user determine which value (0 or nan) to return? For example:

a = np.array([[np.nan, np.nan], [1,np.nan]])

np.nansum(a, axis=1)    # returns array([0, 1])

# This is what I want
np.nansum(a, axis=1, keep_nan=True)    # return array([nan, 1])

The reason that I need this is because when I process image data, I need to distinguish pixels with value 0 and value nan. "Zero" means the pixel has a meaningful value, which is 0; "nan" means the pixel has a meaningless value.

I noticed that there are alrealy some disscussions about this (#6549), i.e. whether should return 0 or nan. Since it is controversial, I suggest given users an option to determine which one they want.

[LeonaTaric]

Now, you can do like this

>>> np.where(np.all(np.isnan(a), axis=1), np.nan, np.nansum(a, axis=1))
array([nan,  1.])

And I think it's a good idea to add a boolean flag to choose.

[seberg]

I am not too deep into the nan functions. From a general perspective, for sum etc. (the non-NaN versions), what would make most sense to me would be default= which would complement/replace initial= with a slightly different meaning.

default= would only be used if there are no values along the axes. It would be tricky to implement it for where=, though. On first sight such an addition seems fine, but possibly very niche.
For example mean() naturally "distinguishes" since it divides by the number of non-NaN values.