SavedModelBundle leak fix

[Craigacp]

This PR adds the deallocators to the TF_Graph and TF_Session that live inside a SavedModelBundle as previously they didn't free their memory on SavedModelBundle.close().

We may want to backport this to the 0.3.x branch. I think the main conflict there is that there are formatting changes in master not present in r0.3 but we could just write a similar patch, it's only 5 lines.

cc @saudet

[Craigacp]

Turns out Graph and Session use different idioms for closing their respective native object, and this difference broke the first version of this fix. We should probably make them use the same idiom, either both using JavaCPP's deallocators, or both directly calling the relevant TF C API functions on close. The split is confusing.

[saudet]

We probably should call TF_Graph.newGraph() here, but I'm not sure...

[Craigacp]

I tried that initially, but Graph.close() doesn't use handle.releaseReference() like Session.close() does, and so it doesn't call the TF_Graph deallocator. I think you're the last person to touch both Graph.delete and Session.delete and I wasn't sure if they were different for a reason.

As I said in the comment above, I think they should be made consistent, but that's going to be a bigger change than just the one which fixes this leak. For example if we switch Graph over to use handle.releaseReference() and then use TF_Graph.newGraph everywhere it'll mean that the TF_Graph inside Graph is now sensitive to enclosing PointerScopes which it currently isn't.

[saudet]

Is it that the TF_Graph becomes managed by the TF_Session and we shouldn't touch it anymore after that? It sounds like that from the documentation of the C API. In that case I'd put the call to TF_NewGraph() in TF_Session.loadSessionFromSavedModel() and forget about it.

[Craigacp]

No, I think the graph is separate from the session, or at least that's my reading of the doc here - https://github.com/tensorflow/tensorflow/blob/master/tensorflow/c/c_api.h#L1229. The TF_NewSession doc above it suggests the TF_Graph is kept alive by the TF_Session, but it doesn't say that the session owns the graph, so I think we're still responsible for closing it.

[killeent-stripe]

@Craigacp Would definitely a +1 to say that backporting this fix to a 0.3.x release would be much appreciated!