Fix segment_reduce() ignoring tailing 0 segments

[mjwen]

Description

segment_reduce() supports zeor-length segments. But it fails if the zero-length segments is at the end of seglen.

For example,

import dgl
import torch

val = torch.ones(10, 3)
seg = torch.tensor([1, 0, 5, 4, 0])
x = dgl.ops.segment.segment_reduce(seg, val, "sum")

yields a value for x as

tensor([[1., 1., 1.],
        [0., 0., 0.],
        [5., 5., 5.],
        [4., 4., 4.]])

but it should be

tensor([[1., 1., 1.],
        [0., 0., 0.],
        [5., 5., 5.],
        [4., 4., 4.],
        [0., 0., 0.]])

This PR fixes this.

[BarclayII]

Thanks for your fix!