fix: restore DSPy 3.x GEPA skill evolution compatibility

[sunwz1115]

Summary

• Make skill evolution compatible with current DSPy/GEPA budget parameters by selecting max_steps, max_full_evals, or max_metric_calls based on the installed GEPA signature.
• Make skill_fitness_metric accept GEPA's reflective feedback call shape while preserving the existing scalar metric behavior for normal evaluation/MIPRO.
• Store the optimizable skill body in DSPy signature instructions so GEPA mutations can be read back as the evolved SKILL.md body.
• Validate complete reassembled skill candidates, including YAML frontmatter, instead of validating body-only markdown.
• Add regression tests covering GEPA kwargs, reflective metric feedback, validation scope, and instruction mutation extraction.

Test Plan

• pytest tests/skills/test_evolution_pipeline_regressions.py -q
• pytest -q

Notes

This PR does not include generated experiment artifacts such as DSPy cache files, run logs, or evolved output snapshots.