Add GEPA recipe for prompt optimization

[sdan]

Adds GEPA: Reflective Prompt Evolution Can Outperform Reinforcement Learning
using Tinker to do batch sampling to evaluate prompt candidates faster. This takes advantage of putting all the "best" prompts on one clock cycle before calling a teacher to reflect and improve the prompt. Comes with GSM8K and AIME and HotpotQA.

[Tiiiger]

hi @sdan thank you for your contribution! I think this is cool but probably not common enough to be merged into the official cookbook.

if you want to create a fork to talk about your work (like this), I am happy to repost!