Models interface should push and pop default allocator context on eval

[mbartling]

To support multiple allocator contexts in the tensor destructor, model interface should push and pop current allocator pointers to a queue in the default current context else risks failed destruction

[mbartling]

@dboyliao found it

[dboyliao]

Example code to reproduce the error:
https://github.com/dboyliao/PyDataTP_2020/blob/master/4_model_interface/model_interface.cpp

[dboyliao]

As a side note, the main issue is that we need to switch the ram & meta allocator back to those in the user space after model eval.
Since model eval will temporary switch allocators to the ones inside the model, not the ones in user space.

If we forget to switch allocators back, tensor deconstructor will fail.