docs: add RLHF reward hacking evaluation lesson

[AKilalours]

Summary

This PR adds a lightweight RLHF reward-hacking evaluation lesson to the prompt evaluations materials.

Why

Preference optimization can improve model behavior, but it can also create failure modes where models optimize for surface-level reward signals rather than the underlying objective. This lesson introduces evaluation patterns for sycophancy, unsafe helpfulness, overconfidence, and unsupported confidence.

Changes

• Added an RLHF reward-hacking evaluation lesson
• Added preference-pair examples for alignment-sensitive behavior
• Added a scoring rubric for reward-hacking failure modes
• Added a suggested evaluation workflow for regression testing
• Linked the lesson from the prompt evaluations README

Testing

• Documentation-only change
• Checked Markdown formatting locally
• Commit signed and verified

[AKilalours]

Hi maintainers,

I opened this as a lightweight prompt-evaluations lesson focused on RLHF reward-hacking failure modes, including sycophancy, unsafe helpfulness, overconfidence, and unsupported confidence.

Happy to revise the wording, placement, or scope if you prefer a different structure for the course materials.