Support for negative segment sizes

In our segmented reduce we state:
> If `d_end_offsets[i] - 1 <= d_begin_offsets[i]`, the *i*^th segment is considered empty.
In our segmented sort and segmented radix sort we state the same thing.

While in the segmented reduce implementation, we check that:
```cpp
if (segment_begin == segment_end)
```
to recognize empty segments. I don't think we test for negative segment sizes.

I suggest we decide if we want to support negative offsets and then add necessary tests / fixes, or we change the docs.