Test failures using PyTorch master

[neerajprad]

While most test failures on pytorch-0.5.0 are resolved, the following tests are still failing with the latest master branch. Some of these are recent failures.

CPU Tests:

• Add naive_ubersum reference implementation + tests #1417 tests/ops/test_contract.py::test_ubersum_sizes (test_contract failing on PyTorch master #1414)
• Fix TraceEnum_ELBO jit tests, etc. #1590 test_svi_irregular_batch_size[JitTraceEnum_ELBO]
• many HMC/NUTS tests with step size adaptation (trace's log_prob of models with Gamma/Dirichlet distributions incompatible with JIT #1487) wontfix - this is an upstream pytorch issue.
• tests/contrib/oed/test_ewma.py::test_ewma - Assertion failures (Support CUDA tensors in OED examples. #1373)
• test_eig_lm[model5-design5-y-w-donsker_varadhan_eig-args5-True-0.3] (Support CUDA tensors in OED examples. #1373)
• test_eig_lm[model10-design10-y-w1-donsker_varadhan_eig-args10-True-0.3] (Support CUDA tensors in OED examples. #1373)
• tests/infer/test_enum.py::test_mixture_of_diag_normals (Assertion failures) (test_mixture_of_diag_normals gives wrong loss on linux #1425)

[fehiepsi]

Closed because these tests have been solved in #1617.