Inconsistent elemwise auto-densify rules

Describe the bug

Sometimes element wise operations return a dense result, sometimes they error. It seems to depend on the broadcasting.

To Reproduce

import sparse
import numpy as np

s = sparse.random((20, 10), .5)
dense_mtx = np.random.rand(20, 10)
dense_vec = np.random.rand(10) 

print(type(s + dense_mtx))
# <class 'numpy.ndarray'>

s + dense_vec

---------------------------------------------------------------------------
ValueError                                Traceback (most recent call last)
<ipython-input-1-8f24377c4243> in <module>
      9 # np.ndarray
     10 
---> 11 s + dense_vec

/usr/local/lib/python3.8/site-packages/numpy/lib/mixins.py in func(self, other)
     19         if _disables_array_ufunc(other):
     20             return NotImplemented
---> 21         return ufunc(self, other)
     22     func.__name__ = '__{}__'.format(name)
     23     return func

~/github/sparse/sparse/_sparse_array.py in __array_ufunc__(self, ufunc, method, *inputs, **kwargs)
    303 
    304         if method == "__call__":
--> 305             result = elemwise(ufunc, *inputs, **kwargs)
    306         elif method == "reduce":
    307             result = SparseArray._reduce(ufunc, *inputs, **kwargs)

~/github/sparse/sparse/_umath.py in elemwise(func, *args, **kwargs)
     47     """
     48 
---> 49     return _Elemwise(func, *args, **kwargs).get_result()
     50 
     51 

~/github/sparse/sparse/_umath.py in __init__(self, func, *args, **kwargs)
    460 
    461         self._check_broadcast()
--> 462         self._get_fill_value()
    463 
    464     def get_result(self):

~/github/sparse/sparse/_umath.py in _get_fill_value(self)
    554         equivalent_fv = equivalent(fill_value, fill_value_array).all()
    555         if not equivalent_fv and self.shape != self.ndarray_shape:
--> 556             raise ValueError(
    557                 "Performing a mixed sparse-dense operation that would result in a dense array. "
    558                 "Please make sure that func(sparse_fill_values, ndarrays) is a constant array."

ValueError: Performing a mixed sparse-dense operation that would result in a dense array. Please make sure that func(sparse_fill_values, ndarrays) is a constant array.

Expected behavior

Consistency. I'd like a dense output for this operation (following https://github.com/pydata/sparse/issues/10#issuecomment-356925562), but would be happy with the consistency of:

• Always error
• Always return dense
• Always return sparse

System

• OS and version: macOS 10.15
• sparse version (sparse.__version__) 0.12.0+8.g75125cd
• NumPy version (np.__version__) 1.20.2
• Numba version (numba.__version__) 0.53.1

Reproduce the two operations from the issue, then read sparse/_sparse_array.py around array_ufunc and sparse/_umath.py around elemwise, _check_broadcast, and _get_fill_value. Compare the broadcasting paths and existing behavior before deciding on the consistent result policy. Done means mixed sparse-dense elementwise operations follow one documented, tested outcome.