[WIP] support for OpenAI Responses API and image gen

[arnavsinghvi11]

Supports OpenAI Responses API
This also supports image generation in DSPy through native calls through OpenAI Responses (Note that you can do this through gpt-image-1 or through models like gpt-4.1-nano, gpt-4.1-mini, gpt-4.1, etc. which call gpt-image-1` under the hood)

Example usage of DSPy image gen:

import base64

lm = dspy.LM(model = 'gpt-4.1-nano', model_type = 'responses', tools=[{
        "type": "image_generation",
        "quality": "low", 
        "size": "1024x1024", 
        "background": "opaque"
    }])
   
dspy.settings.configure(lm=lm)

class DogImageSignature(dspy.Signature):
    """Generate an image of a dog based on the given breed."""
    breed: str = dspy.InputField(desc="The breed of the dog")
    image: dspy.Image = dspy.OutputField(desc="A small image of the dog")

generate_dog = dspy.Predict(DogImageSignature)

response = generate_dog(breed="corgi")

img_obj = response.image

if hasattr(img_obj, "url") and img_obj.url.startswith("data:image/"):
    header, b64data = img_obj.url.split(",", 1)
    with open("corgi.png", "wb") as f:
        f.write(base64.b64decode(b64data))
    print("Image saved to corgi.png")

(For further reference on image gen parameters for OpenAI LMs)

[arnavsinghvi11]

this is related to the implicit warning that certain custom fields must be the only OutputField (aka can't have 2 Image OutputFields or 1 Image, 1 ToolCall OutputField) but maybe we can express this better? @chenmoneygithub @TomeHirata

[arnavsinghvi11]

this got lost in an earlier PR, adding back in

[arnavsinghvi11]

all test related changes are for the refactoring of dspy.Image to handle the Responses Image format (the API strictly requires input_image)

[arnavsinghvi11]

not setting all these params triggers pydantic_core._pydantic_core.ValidationError bc LiteLLM doesn't set defaults for them :(