[BUG] Upcast (but never downcast) the result of a binary operation between decimals (Python)

[shwina]

Currently, a binary operation between two decimal types can result in a downcast if the resulting precision can be accommodated by a smaller decimal type:

In [20]: s = cudf.Series([0, 1, 2], dtype=cudf.Decimal64Dtype(scale=2, precision=5))

In [21]: s
Out[21]:
0    0.00
1    1.00
2    2.00
dtype: decimal64

In [22]: s + s
Out[22]:
0    0.00
1    2.00
2    4.00
dtype: decimal32

This downcast may be inefficient, or at worst, dangerous, depending on downstream operations.

We should never downcast, but upcast if necessary, according to the rules summarized here.

[galipremsagar]

@shwina Assigned it to me, I'll open a PR for these changes.