Fix Python tests by adding GEPA compatibility fallback for OpenTrace experimental

[Copilot]

The failing Python tests job was caused by GEPA trainer aliases resolving to raw strings when opto.features.gepa classes are unavailable in OpenTrace experimental. This broke dynamic trainer resolution tests and any task path depending on GEPA-Base.

• Problem summary

  • GEPA-Base, GEPA-UCB, and GEPA-Beam could not consistently resolve to algorithm classes in CI’s OpenTrace environment, causing failures in dynamic trainer and trace-example stub tests.

• Resolver compatibility update

  • Added a compatibility fallback for GEPA aliases to opto.trainer.algorithms.classical_algorithms.ParetobasedPS.
  • Kept existing preferred resolution path (opto.features.gepa.gepa_algorithms) and only use fallback when those classes are not present.

• Behavioral impact

  • GEPA aliases now resolve to a concrete class across both newer and older OpenTrace layouts, preserving existing alias behavior while avoiding string fallback in supported environments.

_GEPA_COMPAT_FALLBACK = (
    "opto.trainer.algorithms.classical_algorithms",
    "ParetobasedPS",
)

if name in {"GEPA-Base", "GEPA-UCB", "GEPA-Beam"}:
    candidates.append(_GEPA_COMPAT_FALLBACK)