adhere to lazy import rules (#806)

classy_vision/optim/adamw_mt.py

@@ -7,6 +7,7 @@
 from typing import Any, Dict, Tuple
 
 import torch.optim
+from torch.optim import _multi_tensor
 
 from . import ClassyOptimizer, register_optimizer
 
@@ -30,7 +31,7 @@ def __init__(
         self._amsgrad = amsgrad
 
     def prepare(self, param_groups) -> None:
-        self.optimizer = torch.optim._multi_tensor.AdamW(
+        self.optimizer = _multi_tensor.AdamW(
             param_groups,
             lr=self._lr,
             betas=self._betas,