Analyze and implement lifetime behavior of `DeviceSpecificArg`, per device op states, etc.

[KateUnger]

lib/runtime/src/task_spec/device_specific_arg.h:

How do we make sure that ptr actually gets deallocated eventually (don't fix this this PR, but add an issue or just think about it as we'll need to solve this in the future)

struct DeviceSpecificArg {

  private:
    T *ptr;
    size_t device_idx;
}

[lockshaw]

In general, we need to ensure proper deallocation for all of the objects during training that exist across iterations and are not explicitly destroyed by the C ffi (such as flexflow_stop_training and any flexflow_<something>_destroy methods)