More flexible TrainableModel

[corbt]

This PR makes a few changes to the API shape, specifically focused on Model and LocalAPI.

It introduces an abstraction where we separate out PolicyModel, which could be any LLM, and a TrainablePolicyModel, which is specifically a policy that can be trained by our system. This will let us log trajectories from both PolicyModel and TrainablePolicyModel in a unified way.

It also adds a new config field to PolicyModel. This is opaque to our system, but is something we can log to wandb and our file system in the future to track hparams associated with each model run. I'm using the config field in the following way:

agent_002 = art.TrainablePolicyModel(
    name="email-agent-002",
    project="email_agent",
    base_model="Qwen/Qwen2.5-14B-Instruct",
    config=ProjectPolicyConfig(
        max_turns=10,
        training_config=TrainingConfig(
            trajectories_per_group=6,
            groups_per_step=1,
            learning_rate=1.2e-5,
            eval_steps=30,
            val_set_size=100,
            training_dataset_size=4000,
            batch_size=8,
            num_epochs=4,
        ),
    ),
)

agent_004 = agent_002.model_copy(deep=True)
assert isinstance(agent_004.config, ProjectPolicyConfig)
agent_004.name = "email-agent-004"
agent_004.config.max_turns = 30

And then within my training and rollout functions, adjusting behavior based on the config above. I find this pattern helps me stay sane by tracking what properties each model was called with, and keeping a record around of old training runs (by keeping the old config in the codebase).