avoid using legacy constructors Tensor.new in contrib.gp

[fehiepsi]

Follow up of #1842. This is ready to review if tests pass.

[fehiepsi]

It seems that Travis is unstable...

[fritzo]

@fehiepsi it appears there was some confusion on the PyTorch side and these are no longer considered legacy constructors pytorch/pytorch#19995 If pyro.contrib.gp does not currently support the jit, then there is no immediate need to avoid these constructors.

I think it's safe to close this PR.

[fehiepsi]

Actually, I just run the benchmark test with jit loss_fn and get the error RuntimeError: Cannot insert a Tensor that requires grad as a constant. Consider making it a parameter or input, or detaching the gradient Tensor: (might be related to this issue). I am not sure what I should do to resolve it right now so jitted contrib.gp might not be supported in the near future. cc @neerajprad

[neerajprad]

@fehiepsi - Please don't worry about it. This is just a minor release to make Pyro compatible with the latest PyTorch release.