[BUG] Inconsistent handling of integer division by zero

[wence-]

Describe the bug

Type promotion to handle division by zero in __floordiv__ or __mod__ is inconsistently applied if the 0 is already a cudf.Scalar, depending on whether the dtypes of the Scalar and the divided Column match.

Steps/Code to reproduce bug

import cudf

s = cudf.Series([1, 2], dtype="int64")
print(s // 0)
# 0    inf
# 1    inf
# dtype: float64
print(s // cudf.Scalar(0, dtype="int32"))
# 0    inf
# 1    inf
# dtype: float64
print(s // cudf.Scalar(0)) # equivalent to cudf.Scalar(0, dtype="int64")
# 0    9223372036854775807
# 1    9223372036854775807
# dtype: int64
# Technically worse than this, since I think we end up 
# with division by zero of signed integer types in libcudf
# which is UB

Expected behavior

I should always get promotion to float64 (does it need to be float64?) and then inf in the output.

Why does this happen?

NumericalColumn._binaryop checks for division by zero if the op is __mod__ or __floordiv__, but doesn't handle the case where other is a cudf.Scalar. NumericalColumn._wrap_binop_normalization hands back a cudf.Scalar if it is passed a Scalar with the same dtype as the column in question. So in this case, we don't check to see if we have a zero and don't promote to float.

[wence-]

Further inconsistency, is that pandas special-cases bool dtypes and raises NotImplementedError or ZeroDivisionError (depending on ???) for division of a bool series by a bool.

And is inconsistent in its handling of division by a scalar boolean and a series with booleans.

[wence-]

Further inconsistency, is that pandas special-cases bool dtypes and raises NotImplementedError or ZeroDivisionError (depending on ???) for division of a bool series by a bool.

And is inconsistent in its handling of division by a scalar boolean and a series with booleans.

Opened pandas-dev/pandas#49699 to ask.