Various improvements for the JIT

[apaszke]

• TensorType now records device information as well
• Fixed a minor bug in CppOp
• Fixed a bug where passing outputs to save_for_backward would break tracing (outputs wouldn't have the trace associated with them)
• Made certain Functions traceable. To complete the most important parts of word_lanuage_model we're still missing support for Indexing with slices (we should probably have our PyTorch custom index op) and SubConstant that I'm working on. I didn't manage to push it yet, because it needs to be implemented as Neg + AddConstant, but the code breaks with an assertion failure in the graph fuser if I don't set the type for Neg. I'll add support for manipulating types from Python and send a fix for this later.

[ezyang]

For the immediate PR, we need a test for SavedVariables tracing fix. We should also bump up the priority of the autograd/nn-like testing framework for tracing and running (but I wouldn't block the PR on this).

[ezyang]

Thanks. Tests need to pass before going in.