Use updated PyTorch 0.4 tensor creation functions

Right now, we do the following kind of things to get new tensors:

bar_randn = foo.new(foo.size(-2), foo.size(-1)).normal_()
bar_zeros = foo.new(size1, size2).zeros_()
bar_inds = foo.new(size1, size2).zeros_().long()

PyTorch 0.4.0 added the following ways to do this instead that I prefer:

bar_randn = torch.randn_like(foo)
bar_zeros = foo.new_zeros(size1, size2)
bar_inds = foo.new_zeros(size1, size2, dtype=torch.long)

We could also just by carrying around a device string instead of a tensor_cls do:

bar = torch.zeros(size1, size2, dtype=torch.long, device=device_str)

[jacobrgardner]

They already have a device property – I assume he means dtype.

[gpleiss]

LazyVariables should also get a device property, and we should get rid of the tensor_cls method.