pytorch

e9ef1fe3 - [PyTorch Mobile] Add continuous build config for xplat/caffe2

[PyTorch Mobile] Add continuous build config for xplat/caffe2 Summary: Currently this folder isn't covered by continuous build and ideally, it should be covered. I've made everything that is actually used build, but there are test failures (commented out). Specifically: ### Build Failures 1. [Resolved] Vulkan stuff doesn't build because codegen doesn't generate files that Vulkan expects. 2. [Resolved] Vulkan relies in Android dev environment being set up, which doesn't exist on sandcastle machines. I think the resolution should be to restrict Vulkan stuff to the ANDROID platform, but will let AshkanAliabadi (who is the expect on all things Vulkan) provide the appropriate resoltion. 3. [Resolved] Older caffe2 stuff didn't have the deps set up correctly for zlib. 4. [Resolved] Some Papaya stuff didn't have the QPL deps set up correctly. 5. [Resolved] Some tests include cuda, which isn't available on xplat PyTorch Mobile. 6. [Resolved] Missing NNPACK dep on platforms other than ANDROID and MACOS. 7. [Resolved] Maskrcnn binary missing header includes. 8. [Resolved] Braces around scalar initializers in Vulkan Tests. 9. [Resolved] Incorrect header `<vulkan/vulkan.h>` and incorrect BUCK glob path to include it - seems like some completely different header was being included by libvulkan-stub. ### Test Failures 1. [Resolved] Memory Leak on exit in multiple (all?) QNNPACK tests. 2. [Unresolved] Lite Trainer test doesn't explicitly specify dep on input `.ptl` file, resulting in the file not being found in the test when the test attempts to open it. 3. [Resolved] Heap Use after free errors in old caffe2 tests. 4. [Resolved] Heap buffer overflow errors in old caffe2 tests. 5. [Unresolved] Something related to an overload of `at::Tensor` accepting C2 Tensor not being found (new PyTorch test I think). Everything marked `[Unresolved]` above results in stuff that is commented out so that it isn't triggered. It is already currently broken, so it doesn't represent a regression - merely an explicit indication of the fact that it's broken. Everything marked `[Resolved]` above means that it was fixed to function as intended based on my understanding of the intent. Test Plan: Sandcastle. Reviewed By: iseeyuan Differential Revision: D25093853 fbshipit-source-id: e0dda4f3d852ef158cd088ae2cfd44019ade1573