Add GEPA prompt optimization tutorial and clean up Qwen3 configs

[darisoy]

Description

This PR introduces a new tutorial notebook demonstrating how to use GEPA (Generic Evaluation and Prompt Adaptation) to optimize system prompts for MaxText models. It also includes minor configuration cleanups for Qwen3 models in the checkpoint conversion utility.

The specific changes include:

• New Tutorial Notebook (src/maxtext/examples/maxtext_with_gepa.ipynb):
  • Provides environment setup instructions for a single TPU GCE VM using uv.
  • Covers downloading and converting the qwen3-4b checkpoint.
  • Demonstrates how to start and serve the model using MaxText/vLLM on a TPU.
  • Walks through using GEPA to optimize the model's system prompt against the AIME dataset.
  • Includes a before-and-after comparison of the model's accuracy to showcase the optimization results.
• Config Cleanup (src/maxtext/checkpoint_conversion/utils/hf_model_configs.py):
  • Removes the use_sliding_window parameter from the Qwen3 configurations, as it is no longer supported.

Tests

The changes primarily consist of an example notebook. The notebook contains reproducible steps to run end-to-end tests on a TPU VM, which covers:

• Checkpoint conversion of Qwen3-4B.
• vLLM serving initialization.
• Executing the GEPA optimization loop and evaluating both the seed prompt and optimized prompt outputs.

Checklist

Before submitting this PR, please make sure (put X in square brackets):

• I have performed a self-review of my code. For an optional AI review, add the gemini-review label.
• I have necessary comments in my code, particularly in hard-to-understand areas.
• I have run end-to-end tests tests and provided workload links above if applicable.
• I have made or will make corresponding changes to the doc if needed, including adding new documentation pages to the relevant Table of Contents (toctree directive) as explained in our documentation.

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.

📢 Thoughts on this report? Let us know!