[BUG] Incorrect broadcasted equality for scalar comparison to int64 columns

[wence-]

Describe the bug

Equality of a NumericalColumn with large int64 values against another large int64 value can wrongly advertise false positives.

Steps/Code to reproduce bug

import cudf
import numpy as np

s = cudf.Series([2**63 - 10, 2**63 - 100], dtype=np.int64)
print(s == np.int64(2**63 - 1))
# 0    True
# 1    True
# dtype: bool

Expected behavior

Equality for int types should be correct.

Why does this happen?

Before heading into libcudf we need to turn the scalar host value into a device scalar, for which a dtype must be deduced. This is done in NumericalColumn.normalize_binop_value.

The relevant code is:

        other_dtype = np.min_scalar_type(other)
        if other_dtype.kind in {"b", "i", "u", "f"}:
            if isinstance(other, cudf.Scalar):
                return other
            other_dtype = np.promote_types(self.dtype, other_dtype)

So first we inspect the value and ask numpy to come up with a dtype that can represent this value, then we promote the two types of the column and the deduced dtype to a common type, cast our target value to that and continue.

What goes wrong?

val = np.int64(2**63 - 10)
mintype = np.min_scalar_type(val) # => np.uint64 (oh no!)
otype = np.promote_types(np.int64, np.uint64) # => float64 (oh no!)

Numpy prefers unsigned over signed types min_scalar_type which is problematic, because if the column type is signed, then we think we can no longer represent the RHS value in the column type. So we upcast to float (lossy) and call into libcudf with binaryop(column, device_scalar_as_double(val)) and now we get standard C++ type promotion on the pointwise equality comparison.

Environment overview (please complete the following information)

cudf.__version__ => '22.12.00a+193.g991c86b13a'