Add PydanticFieldGEPA: Per-Field Description Evolution for Data Extraction

[mango-panther]

closes #12
Adds Pydantic field description evolution as a mode within the existing GEPA optimizer. When a pydantic_model is provided, GEPA treats each field description as a separate evolvable component instead of evolving predictor instructions.

Usage:

from pydantic import BaseModel, Field                                                                                                    
import dspy                                                                                                                              
                                                                                                                                         
class Contact(BaseModel):                                                                                                                
    name: str = Field(description="Contact's full name")                                                                                 
    email: str = Field(description="Email address")                                                                                      
                                                                                                                                         
optimizer = dspy.GEPA(                                                                                                                   
    metric=extraction_metric,                                                                                                            
    pydantic_model=Contact,                                                                                                              
    evolvable_fields="all",                                                                                                              
    reflection_lm=dspy.LM("openai/gpt-4o-mini"),                                                                                         
    max_metric_calls=500,                                                                                                                
)                                                                                                                                        
                                                                                                                                         
optimized = optimizer.compile(extractor, trainset=examples)      

New GEPA Parameters:

• pydantic_model: Pydantic model defining extraction schema (triggers pydantic mode)
• evolvable_fields: "all", list of field names, or None for seed_prompt only
• field_scorers: Optional per-field scorer functions
• field_weights: Optional weights for field score aggregation
• input_field_name, output_field_name, base_instruction: Extraction configuration

[mango-panther]

@TomeHirata Could you also review this one when you get time

[TomeHirata]

This shouldn't be necessary if we import them lazily where needed. Where does the circular import occur?

[mango-panther]

1. dspy/teleprompt/init.py imports from .gepa.gepa import GEPA
2. This triggers loading of dspy/teleprompt/gepa/init.py first (package init)
3. With eager imports, it loads pydantic_field which uses dspy.LM in type hints
4. But dspy isn't fully loaded yet → circular import

[TomeHirata]

Can we make the Pydantic field evolution another mode of the existing GEPA optimiser instead of introducing PydanticFieldGEPA? We can add a new argument to GEPA and switch the GEPA adapter class used internally based on the argument.

[mango-panther]

Can we make the Pydantic field evolution another mode of the existing GEPA optimiser instead of introducing PydanticFieldGEPA? We can add a new argument to GEPA and switch the GEPA adapter class used internally based on the argument.

Made these change. I have also upated PR description based on this new implementation.

[mango-panther]

@TomeHirata, can you review this again? I have incorporated all suggested changes.