[FEA] Provide a way to specify the maximum allowable precision for integers/floats

**Is your feature request related to a problem? Please describe.**

GPU memory is a valuable resource, and using int64/float64 columns where int32/float32 would suffice means using 2x as much memory unnecessarily. As opposed to scientific computing, 32-bit data types (or lower) are sufficient for many data science applications.

Even only 32-bit data types as inputs, the resulting output can be a 64-bit type:

```python
>>> cudf.Series([1, 2, 3], dtype="int32") + cudf.Scalar(1, dtype="float32")
0 2.0
1 3.0
2 4.0
dtype: float64
```

(this is consistent with Pandas and NumPy)

**Describe the solution you'd like**

It would be nice to be able to specify a maximum bitwidth for integer/floating types. If an operation would result in a value greater than could be accommodated, simply overflowing would be acceptable.

This could be another use case for [cudf.config](https://github.com/rapidsai/cudf/issues/5311).

**Describe alternatives you've considered**

The user can carefully cast results back from 64bit to 32bit to reduce memory usage, but this is tedious and does not help with peak memory usage.