torch.memory_format are not singletons

[ezyang]

🐛 Describe the bug

Memory format retrieval allocates fresh memory format with THPMemoryFormat_New

tools/autograd/templates/python_nn_functions.cpp:    PyTuple_SET_ITEM(tuple.get(), 3, THPMemoryFormat_New(opt_memory_format.value(), "unused_name"));
torch/csrc/MemoryFormat.cpp:PyObject *THPMemoryFormat_New(at::MemoryFormat memory_format, const std::string& name)
torch/csrc/MemoryFormat.h:PyObject * THPMemoryFormat_New(at::MemoryFormat memory_format, const std::string& name);
torch/csrc/utils/tensor_memoryformats.cpp:    PyObject* memory_format = THPMemoryFormat_New(format, module_name + name); \

Versions

master

cc @VitalyFedyunin @jamesr66a