Add group_batch functionality

[tzuhanchang]

Tensor src is batched into groups according to index tensor in the given dimension dim.
Padding is applied to enforce consistent size of batched groups.

Examples:

>>> src = tensor([[0.1349, 0.8266],
...               [0.3651, 0.1737],
...               [0.0211, 0.7000],
...               [0.1971, 0.2903],
...               [0.4086, 0.7221]])
>>> index = torch.LongTensor([0,0,1,2,2])
>>> group_batch(src, index, dim=0, pad_size=3, return_mask=True)
(tensor([[[0.1349, 0.8266],
          [0.3651, 0.1737],
          [  -inf,   -inf]],
         [[0.0211, 0.7000],
          [  -inf,   -inf],
          [  -inf,   -inf]],
         [[0.1971, 0.2903],
          [0.4086, 0.7221],
          [  -inf,   -inf]]])
tensor([[[ True,  True],
         [ True,  True],
         [False, False]],
        [[ True,  True],
         [False, False],
         [False, False]],
        [[ True,  True],
         [ True,  True],
         [False, False]]]))