Ndarray sum keep_dims:true does not always keep dimensions

[mooreryan]

I'm not sure if this was a design decision or a bug, but when you run the sum function with no axis argument, and keep_dims:true, the results are not what I would expect given the way keep_dims works in other cases, as well as given how the keepdims argument in numpy's sum function works.

# #require "owl, owl-top";;
# module A = Owl.Dense.Ndarray.D;;
# let a = A.of_arrays [| [| 1.; 2.; 3. |];
                         [| 4.; 5.; 6. |] |];;
val a : A.arr =
   C0 C1 C2
R0  1  2  3
R1  4  5  6

# A.sum ~keep_dims:true a;;
- : A.arr =
  C0
R 21

# A.sum ~keep_dims:true a |> A.shape;;
- : int array = [|1|]

Given the way the keep_dims:true arg works when providing dimensions, (and the way numpy) works, I would expect this result:

# A.sum ~keep_dims:true a;;
- : A.arr =
   C0
R0 21

# A.sum ~keep_dims:true a |> A.shape;;
- : int array = [|1, 1|]

---

Here's a summary of the function's current behavior

in shape	axis	keep_dims	out shape
2, 3	NA	false	1
2, 3	NA	true	1*
2, 3	0	false	3
2, 3	0	true	1, 3
2, 3	1	false	2
2, 3	1	true	2,1

This makes sense except for the case I showed above (the row with the star in this little table).

---

And here is the whole utop session if you want to view it:

sum.md