Set up proper models in eval

[allenzren]

Previously, the eval model loads the network from the saved checkpoint during fine-tuning, which is the frozen base policy.

Now set up a dedicated eval model class that loads the actor and actor_ft weights from the saved checkpoint. actor is for the frozen pre-trained policy, and actor_ft is for the fine-tuned policy

For eval the config needs to have the correct ft_denoising_steps specified from fine-tuning, such that the eval model knows which model (actor or actor_ft) to use depending on the current denoising step

#29