feat(gepa): GEPA/GEPA-Flow Pareto optimizers + docs alignment

[monotykamary]

• What kind of change does this PR introduce? (Bug fix, feature, docs update, ...)

  • Feature + Docs: Implement and document GEPA (single-module) and GEPA-Flow (multi-module) Pareto optimizers; migrate docs to the new compile-based multi-objective flow.

• What is the current behavior? (You can also link to an open issue here)

  • No GEPA optimizer existed prior; only MiPROv2
  • Tracking: Resolves Add GEPA Optimizer #316

• What is the new behavior (if this is a feature change)?

  • Adds GEPA and GEPA-Flow optimizers using compile for multi-objective optimization (Pareto sampling, per-instance selection, hypervolume reporting).
  • Introduces options (e.g., maxMetricCalls, tieEpsilon, paretoMetricKey/paretoScalarize) and strict acceptance aligned with the GEPA paper and repo.
  • Updates docs/OPTIMIZE.md to provide GEPA/GEPA-Flow examples, result interpretation, and performance notes.

• Other information:

  • References:

    • Paper: https://arxiv.org/pdf/2507.19457
    • Reference code: https://github.com/gepa-ai/gepa

  • Example runs (for reproducibility):

    GEPA example (src/examples/gepa.ts)

    $ npm run tsx src/examples/gepa.ts
    
    > @ax-llm/ax-monorepo@14.0.20 tsx
    > node --env-file=.env --import=tsx src/examples/gepa.ts
    
    🚀 Running GEPA Pareto optimization (accuracy + brevity)...
    
    ● Optimization Started
    ──────────────────────────────────────────────────
      Optimizer: GEPA
      Examples: 11 training, 6 validation
      Config: {"numTrials":20,"minibatch":true}
    ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
    
    ● Round 1/20
      Score: 16.200 (best: 16.200)
      Config: instructionLen=484.00, parent=0.00, totalRounds=20.00
    
    ● Round 2/20
      Score: 16.200 (best: 16.200)
      Config: instructionLen=437.00, parent=0.00, totalRounds=20.00
    
    
    ✅ Pareto optimization complete
    Front size: 1
    Hypervolume (2D): 0.19999999999999998
    
    Top Pareto points:
      #1: accuracy=0.500, brevity=0.400, config={"candidate":0}
    
    🎯 Chosen compromise score (0.7*acc + 0.3*brev): 0.470
    Chosen configuration: {"candidate":0}
    

    GEPA-Flow example (src/examples/gepa-flow.ts)

    $ npm run tsx src/examples/gepa-flow.ts
    
    > @ax-llm/ax-monorepo@14.0.20 tsx
    > node --env-file=.env --import=tsx src/examples/gepa-flow.ts
    
    [AxFlow] new AxFlow() is deprecated. Use flow() factory instead.
    🚀 Running GEPA-Flow Pareto optimization (accuracy + brevity)...
    
    ● Optimization Started
    ──────────────────────────────────────────────────
      Optimizer: GEPA-Flow
      Examples: 11 training, 6 validation
      Config: {"numTrials":16,"minibatch":true}
    ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
    
    ● Round 1/16
      Score: 6.000 (best: 6.000)
      Config: modules=2.00, totalRounds=16.00
    
    ● Round 2/16
      Score: 6.850 (best: 6.850)
      Config: modules=2.00, totalRounds=16.00
    
    ● Round 3/16
      Score: 5.700 (best: 5.700)
      Config: modules=2.00, totalRounds=16.00
    
    ● Round 4/16
      Score: 6.300 (best: 6.300)
      Config: modules=2.00, totalRounds=16.00
    
    ● Round 5/16
      Score: 5.850 (best: 5.850)
      Config: modules=2.00, totalRounds=16.00
    
    ● Round 6/16
      Score: 6.500 (best: 6.500)
      Config: modules=2.00, totalRounds=16.00
    
    
    ✅ Pareto optimization complete
    Front size: 2
    Hypervolume (2D): 0.3
    
    Top Pareto points:
      #1: accuracy=0.667, brevity=0.450, config={"candidate":0}
      #2: accuracy=0.667, brevity=0.450, config={"candidate":1}
    

[dosco]

sorry for the delay, was travling, will take a look later today.