Refactor cpp tests

[ailzhang]

• Refactor all uses of at:: with torch:: APIs in test_aten_xla_tensor.cpp, basically this file shouldn’t touch any at:: and torch_xla::/bridge internals, instead only torch:: APIs are allowed there. Other files are used to test internals specifically, so at::/torch_xla/bridge are fine there.
• Refactor a few test helper functions accordingly, make sure all comparison happens on CPU tensor.
• Refactor test CMakeList to rely include Torch package instead of linking each .so file separately. (hard to maintain).
• Remove torch_xla::ToTensor(at::Tensor) which was used to extract Tensor out of a Variable. Upstream was fixed by Use at::AutoNonVariableTypeMode before calling ATen tensor factory functions pytorch#22364 so that we make sure all tensors passed to xla has tensor.is_variable()=false. We can safely remove this hack now.
• After this PR, when a new op lowering is added, we should add tests in test_aten_xla_tensor.cpp and test_tensor.cpp for it.
  Note that this PR is not clang-formatted yet, we need to find a open-source based clang-format config :P (otherwise it's not convenient to ask every OSS contributor to have clang-format-9 installed from source).
  Detailed context: https://docs.google.com/document/d/1nkiPPhcQy6RpBQ13Y-HGkuefnY-B-3kgFveQ0h85Jjo/edit#

[asuhan]

Do we still need this method? I'd try to remove it.

[dlibenzi]

Please don't, test_tensor.cpp still uses it.

[dlibenzi]

@ailzhang Please, the next time send the document first, let's discuss it, and the you do the code ;)
With a 6+K LOC PR, most of which a search&replace, but also bundles other stuff, it is better to split the big search&replace part into one PR (which can be LGTM-ed in one second), and the "other stuff" which requires more thinking into a separate one.
This prevents the reviewers in going through 6K LOC picking up search&replace parts, for the things which need more attention.

[dlibenzi]

Please don't, test_tensor.cpp still uses it.

[dlibenzi]

No need to have different name, ForEachDevice() should be fine.

[ailzhang]

FYI I've run this PR manually against the bert model and it stabilizes at 3 graphs.

[dlibenzi]

Approval on the basis CPU and TPU tests are green with this.