[FEA] Extend cudf's replace API `normalize_nans_and_zeros` to accept nested types

Currently, the replace API `normalize_nans_and_zeros` only accepts a non-nested input column that is either FLOAT32 or FLOAT64 type. As a result, when we want to replace `-NaN` or `-0.0` in a child column of a nested column, we have to manually recursively iterate over all the children columns of the given column, check to see whether the current child column is FLOAT32 or FLOAT64 type before calling to `normalize_nans_and_zeros`.

We should extend `normalize_nans_and_zeros` to do this automatically. In other words, it should accept an input column of any type.