[BUG] `IndexedFrame._split` is inconsistent for empty dataframes

This came up while reviewing #12704.

Quoth the documentation:

> Split a frame with split points in ``splits``. Returns a list of Frames of length `len(splits) + 1`.

Which is true, except if the input dataframe is empty:

```python
import cudf
df = cudf.DataFrame({"a": []})
print(df._split([0])) # => []
```

This makes writing generic code difficult, since we're expecting to get back a list of N+1 things to iterate over, but in this case we don't.

Slicing empty dataframes works fine (and reproduces semantically what you "expect" from slicing empty python lists):

```python
df[:0], df[0:] # => (Empty DataFrame, Empty DataFrame)
```

(Arguably slicing with an out of bounds index should raise an `IndexError`, but that ship has sailed.)

What I would like:

```python
splits = [...]

assert df._split(splits) == [df[s:e] for s, e in zip([None] + splits, splits + [None])]
```