adhere to lazy import rules (#806)

Summary:
Pull Request resolved: #806

Lazy import changes `Python` import semantics, specifically when it comes to initialization of packages/modules: https://www.internalfb.com/intern/wiki/Python/Cinder/Onboarding/Tutorial/Lazy_Imports/Troubleshooting/

For example, this pattern is not guaranteed to work:

```
import torch.optim
...
torch.optim._multi_tensor.Adam   # may fail to resolve _multi_tensor
```

And this is guaranteed to work:

```
import torch.optim._multi_tensor
...
torch.optim._multi_tensor.Adam   # will always work
```

A recent change to `PyTorch` changed module initialization logic in a way that exposed this issue.

But the code has been working for years? This is the nature of undefined behavior, any change in the environment (in this the `PyTorch` code base can make it fail.

Differential Revision: D58881291

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,