Log metrics horizontally in W&B to simplify comparison to future runs

[arcticfly]

Utility function to simplify comparison of constant model's metrics against those of newly trained models.

In the screenshot below, the golden line is the error rate of a comparison model.

Example:

await log_constant_metrics_wandb(
    model=model,
    num_steps=num_horizontal_run_steps,
    split_metrics={
        "val": metrics_dict,
        "val_failures": {
            k: v for k, v in metrics_dict.items() if "failed" in k
        },
    },
)