Fix missing Weights & Biases metric logging for tracking data

[marina-aoyama]

📝 Description
This PR fixes an issue where certain training metrics (e.g., rewards) were not appearing in Weights & Biases, even when sync_tensorboard=True was enabled.

🔧 Problem
Although W&B was configured to sync TensorBoard logs, some metrics stored in self.tracking_data were not consistently propagated to W&B dashboards. This resulted in incomplete experiment tracking, particularly for key RL metrics such as episode rewards.

💡 Root cause
W&B TensorBoard synchronization does not reliably capture all scalar logs in this setup, especially when metrics are updated outside of immediate TensorBoard flush cycles.

✅ Solution

• Add explicit W&B logging for self.tracking_data via wandb.log
• Ensure logging only occurs when 1) W&B is enabled in config, and 2) wandb is installed.

✨ Change summary

• Added safe import handling for wandb
• Added runtime check for wandb.run
• Added explicit metric logging alongside TensorBoard logging

if self.cfg.experiment.wandb and wandb.run is not None:
    wandb.log({k: float(np.mean(v))}, step=timestep)

📊 Impact

• Restores missing metrics in W&B dashboards (including rewards)
• Keeps TensorBoard logging unchanged
• Maintains backward compatibility when W&B is disabled or not installed
• Improves experiment reproducibility and monitoring reliability