nn.EmbeddingBag doesn't seem to support None offset for 2D tensors

Example
```
import torch
from torch import nn
from torch.autograd import Variable
a = Variable(torch.LongTensor([[1, 0, 2], [2, 1, 0], [0, 0, 1]]))
embed = nn.EmbeddingBag(3, 4)
embed(a)
```
Gives
RuntimeError: expected a Variable argument, but got torch.LongTensor