feat: move trajectory logging from backend to frontend

[corbt]

Summary

Implements RFC #511 - moves trajectory persistence and metrics reporting from the backend to the Model class, allowing backends to focus solely on training.

Key Changes

Model Class Additions

• Model.log() now handles parquet writing, metrics calculation, history.jsonl, and wandb logging inline
• New base_path attribute (default: ".art") - configurable output path
• New report_metrics attribute - explicit control over wandb logging (None = auto-detect, ["wandb"] = enable, [] = disable)
• New _get_wandb_run() method for lazy wandb initialization
• get_step() method now available on base Model class (returns 0 for non-trainable)

TrainableModel Changes

• train() now calls log() first, then delegates to backend
• delete_checkpoints() reads history.jsonl locally to determine best checkpoint, then calls backend

Backend Simplification

• Removed Backend._log() method entirely
• Renamed _delete_checkpoints() to _delete_checkpoint_files(steps_to_keep)
• Removed _log_metrics(), _get_wandb_run(), _get_reward_std_dev_learning_rate_multiplier() from LocalBackend
• Backends now focus only on training and checkpoint file deletion

Breaking Changes

For Backend Interface (internal API, not user-facing)

1. Backend._log() method removed
2. Backend._delete_checkpoints(benchmark, smoothing) renamed to _delete_checkpoint_files(steps_to_keep)

For ServerlessBackend Users

• Trajectories are now saved locally (via Model.log()) instead of being sent to the API
• This is a functional change but the user API remains the same

Removed Feature

• scale_learning_rate_by_reward_std_dev internal config option removed - users who need this can implement it themselves using the history.jsonl data

Migration Guide

No changes required for standard usage. The following APIs work exactly as before:

await model.log(trajectories, split="val")
await model.train(trajectory_groups, config=art.TrainConfig(learning_rate=5e-5))
await model.delete_checkpoints()

New optional configuration:

model = art.TrainableModel(
    ...,
    base_path="/custom/path",      # Custom output directory (default: ".art")
    report_metrics=["wandb"],      # Explicit wandb control (default: None = auto)
)

Tests

Added comprehensive test suite in tests/unit/test_frontend_logging.py covering:

• Parquet compatibility with existing readers
• history.jsonl format compatibility with polars
• Path structure matching LocalBackend locations
• Metric calculation and prefixing
• Wandb integration logic
• New model attributes

Closes #511