BUG: unique with NaNs and along an axis inconsistent with flat version

[dstansby]

Describe the issue:

Running np.unique on a flat array with multiple NaNs only returns one NaN. Running along an axis that has duplicated rows containing NaNs returns both rows however.

I suspect this is related to #20326 - whatever the decision here I think the result should be consistent between axis=None and axis=0.

Reproduce the code example:

import numpy as np

print(np.unique([np.nan, np.nan, 2.]))
# [ 2. nan]

a = np.array([[np.nan, 2.],
              [np.nan, 2.]])
print(np.unique(a))
# [ 2. nan]
print(np.unique(a, axis=0))
# [[nan  2.]
#  [nan  2.]]
#
# Expected:
# [[nan  2.]]

Error message:

No response

NumPy/Python version information:

1.22.1
3.9.7 (default, Sep 16 2021, 08:50:36)
[Clang 10.0.0 ]

[rjeb]

Reproduced this behavior on Python 3.9.7 and the current numpy build.

The inconsistent behavior with the "axis" keyword was mentioned in the original issue #2111 but an implementation for NaN-equals-Nan seems to not have been made for it yet.