Implement debug mode and improve context and metadata collection

[Yahiewi]

This PR brings 3 majors changes:
-Context collection now works as agreed, by traversing backwards from the code cell to the nearest markdown cell and gathering it and all intermediate cells.
-Added a debug mode which logs the prompts and answers under /tmp/jupyter-ai-tutor. This can be activated by running jupyter lab --JupyterAITutorApp.debug=True.
-Implemented ROT13 decoding for the reference_solution found in the cell's metadata, as well as formatting for the evaluation criteria from the metadata into more readable prose for the LLM.
-Added metadata examples for the matplotlib_exercices.ipynb example to test these changes.
-The payload sent to the LLM now follows this structure:

=== SYSTEM PROMPT === 
This is the prompt from `AGENT.md`
=====================
=== USER MESSAGE ===
<student_context> The context related to the code cell from which the 'help' command was ran. <student_context> 
<student_answer> The active cell (which the student might have changed) <student_answer>
<reference_solution> Decoded using the ROT13 decoder from the cell metadata <reference_solution> 
<evaluation_criteria> Formatted from the cell metadata <evaluation_criteria>
====================

Screencast.from.2026-07-03.14-14-52.mp4

In this screencast, we inspect the two test files so we can see what exactly was the prompt sent to the LLM and the answer it generated.

[github-actions]

👈 Launch a Binder on branch Yahiewi/jupyter-ai-tutor/debug