Scalar numpy types handled incorrectly in fill method

[HDembinski]

The fill method does not broadcast correctly in all cases.

import boost_histogram as bh
import numpy as np

r = bh.axis.Regular(5, 0, 1)
h = bh.Histogram(r, r)
h.fill(1, [1, 2]) # ok
h.fill(np.float64(1), [1, 2]) # fails

It not only does not work, but the error message is also very confusing:
ValueError: spans must have compatible lengths

[henryiii]

A few notes before a proper fix (tomorrow), the issue is in _fill_cast:

• np.ascontiguousarray turns scalars into 1 element arrays.
• np.isscalar might be useful here
• np.asarray seems to do the right thing, might be better (pybind11 forces contiguous already, and there if there is a copy, NumPy should make it contiguous by default, so it should be still impossible to make two copies. In fact, due to the limited number of datatypes we support, this might save a copy...

[patrickbryant]

I have found a similar broadcasting issue when filling a StrCategory histogram with an array of weights:

import boost_histogram as bh
r = bh.axis.StrCategory([])
h = bh.Histogram(r, r, storage=bh.storage.Weight())
h.fill('test','this',weight=1) # ok
h.fill('test',['this','error'],weight=[1,2]) # ok
h.fill('test','this',weight=[1,2]) # fails with ValueError: spans must have compatible lengths

This is with boost_histogram version 1.3.2