feat: add notebook explaining ChatCompletion to SFT transformation

[jalateras]

Summary

• Added comprehensive documentation explaining how ChatCompletion training data is transformed internally
• Created new notebook with step-by-step examples and visualizations
• Addresses the knowledge gap about OpenAI's internal SFT processing pipeline

What Changed

Created a new Jupyter notebook Understanding_ChatCompletion_SFT_transformation.ipynb that explains:

• How structured ChatCompletion messages are converted to linear sequences
• The role of special tokens and message boundaries
• Tokenization process and token ID conversion
• Loss mask creation and why only assistant tokens contribute to training loss
• Gradient flow during backpropagation
• Practical implications for designing effective training data

Why This Matters

Users frequently ask about what happens "under the hood" when they provide training data in ChatCompletion format for fine-tuning. This documentation fills that gap by providing a detailed technical explanation with code examples that demonstrate each transformation step.

Testing

The notebook includes runnable code examples that:

• Demonstrate the transformation pipeline
• Show tokenization in action
• Explain loss calculation with concrete examples
• Provide visualization of the training process

Fixes #2075