Update in inference.py -- omitting the optimizer's state

[vhiwase]

During the inference phase when loading a trained model, there's a focus on exclusively loading the model's weights while disregarding or not loading the optimizer's state. This adjustment can be particularly beneficial when utilizing a pre-trained model for tasks such as inference or transfer learning, where the optimizer's state information might not be necessary.

[vhiwase]

@microsoft-github-policy-service agree