Incorrect output of cupy.logaddexp()

[leofang]

For this case, mathematically we should get inf, but CuPy returns nan:

>>> np.logaddexp(np.inf, np.inf)
inf
>>>
>>> cp.logaddexp(np.inf, np.inf)
array(nan)

The reason is in0-in1 gives nan when both are inf, and it propagates all the way out:

cupy/cupy/_math/explog.py

Line 73 in 4469fae

'out0 = fmax(in0, in1) + log1p(exp(-fabs(in0 - in1)))',

[leofang]

Note: this relates to the test_indexing.py failure seen in #5698 (comment).