feat: add configurable multi-step Best-of-N evolution

[Enoch-H-Kang]

Description

This PR lets GEPA try several possible prompt/text improvements at each reflection step, pick the best-looking one, and optionally repeat that process multiple times before evaluating the final candidate.

GEPA generally produce one textual evolution per component per reflection step. With this change, users can configure:

• bon: how many candidate textual evolutions to generate per step
• itr: how many sequential inner evolution steps to run before evaluation
  For example, bon=3, itr=2 means GEPA generates 3 possible updates, asks the reflection LM to choose the best one, then uses that selected update as the starting point for another round of 3 updates.

What Changed

• Added bon and itr to gepa.optimize(...).
• Added bon and itr to optimize_anything.ReflectionConfig.
• Implemented multi-step Best-of-N textual evolution in the modern StatelessReflectionLM path.
• Preserved support for the standalone DSPy adapter proposal path.
• Added metadata for inner textual-gradient iterations.
• Added validation for invalid bon / itr values.
• Added tests covering Best-of-N selection, chained inner iterations,
  config forwarding, and validation.

Validation

Ran the repository checks from CONTRIBUTING.md:

• uv sync --extra dev --python 3.11
• uv run pre-commit run --files ...
• uv run pytest tests/
  • 564 passed, 4 skipped
• uv run pyright
  • 0 errors, 0 warnings

[semanticdiff-com]

Review changes with  

Changed Files

File	Status
src/gepa/proposer/reflective_mutation/reflective_mutation.py	60% smaller
src/gepa/optimize_anything.py	47% smaller
tests/test_reflection_lm.py	30% smaller
src/gepa/proposer/reflective_mutation/reflection_lm.py	19% smaller
src/gepa/api.py	17% smaller
src/gepa/adapters/dspy_adapter/dspy_adapter.py	0% smaller
tests/test_optimize.py	0% smaller