Printing large tensors is slow in master.

>>> x = torch.randn(1000, 1000, 1000)
>>> %timeit repr(x)
1 loop, best of 3: 21.8 s per loop

This is important because we often deal with very large tensors.

Note that printing large tensors is much faster in NumPy (10,000x in this case):

>>> %timeit repr(x.numpy())
100 loops, best of 3: 2.18 ms per loop