[Feature Request] Documentation about recommended use of .train() and .eval()

[valterschutz]

Motivation

The ClipPPOLoss documentation says "it is highly recommended to keep your model in eval mode" and there is a ConsistentDropoutModule which makes me think that having dropout in normal nn.Modules is not recommended?

I just happened to stumble across this information in the docs, it is not mentioned in any of the tutorials that I have read.

Solution

• Add a section in the Getting started with model optimization page that explains whether .eval() should be used consistently or not.
• Use .eval() consistently in the tutorials, notably the PPO one.

Checklist

• I have checked that there is no similar issue in the repo (required)