[Feature Request] Make `SavedTensorDict` lazy

Nearly all ops in SavedTensorDict could be done lazily:

class SavedTensorDict(_TensorDict):
    def register_op(self, op, args, kwargs):
         self._op_register.append((op, args, kwargs))

    def some_op(self, ..):
        self.register_op('some_op', ...)

    def _load(self):
        td = load(..)
        for op, args, kwarg in self._op_register:
            td = getattr(td, op)(*args, **kwargs)
        return td
    
    def _save_td(self, td):
        ....
        self_op_register = []