Logging Support for RL Training

This pull request introduces logging functionality to provide deeper insights into Reinforcement Learning (RL) training.

Features

• Implements a general, easily extensible logger that can be subclassed to support different backends (e.g., Weights & Biases).
• Currently supports TensorBoard logger (TB_Logger).
• Fully backward compatible — older code will continue to work, but logging can now be optionally enabled.
• Comes with example script: examples/14_hello_world_logging.py, which demonstrates logging on top of what examples/01_run_hello_world.py already provides.

Logged Metrics

The logger tracks important RL metrics, including:

• avg_sliding_window
• advantage
• clip_frac
• entropy
• KL
• grad_norm
• loss
• policy_log_prob
• policy_ratio
• Maximum token length fed in during a batch
• Maximum number of unmasked completion tokens during a training batch

Visual Demonstration

The images below showcase the logged metrics. Running examples/14_hello_world_logging.py does not break training, as evidenced by the reward/avg_sliding_window metric converging to 1.0.