Right now this just duplicates #5126 - so make sure to address the comments there

My new changes pass through all scalars, and then for arrays of size greater than 1 you get a ValueError of "Array needs to be of size 1".

This changed the behavior of the following:

>>> type(np.asscalar(np.float64(1)))
numpy.float64  # previously float
>>> type(np.asscalar(np.array(1.0)))
float  # as before

This function strikes me as one in the "mistake we're now stuck with" category, which might be best to just leave alone. Calling item() in general is a pretty bad idea.