Add EvilTwin optimizer for evil twin prompt optimization

[ramisbahi]

This PR introduces the EvilTwin optimizer to DSPy, implementing the Greedy Coordinate Gradient (GCG) algorithm for evil twin prompt optimization. EvilTwin generates evil twin prompts that induce similar model outputs while appearing garbled or obfuscated, based on the "Prompts have evil twins" paper.

Key Features:

• Uses KL divergence minimization to iteratively modify prompts to achieve a similar output distribution.
• Runs on local models (default: "EleutherAI/gpt-neo-125M") since it requires gradients, logits, and token-level likelihoods, which API-based LLMs don’t expose.
• Supports customizable optimization settings (e.g., n_epochs, batch_size, top_k, gamma for fluency penalty).
• Provides an easy way to retrieve the final optimized prompt via optimizer.optimized_prompt.

Example Usage:

from dspy.teleprompt.evil_twin import EvilTwin

predictor = dspy.Predict('question -> answer')
q = "Describe the definition of artificial intelligence in one sentence."

optimizer = EvilTwin(question=q)
optimized_predictor = optimizer.compile(program=predictor)

print("Optimized Evil Twin Prompt:", optimizer.optimized_prompt)
original_response = predictor(question=q)
evil_twin_response = optimized_predictor(question=q)

print("Original Output:", original_response.answer)
print("Evil Twin Output:", evil_twin_response.answer)

Notes:

• EvilTwin is best run on a GPU due to the computational cost of token gradient updates.
• Future work may include warm start initialization, as proposed in the Evil Twins paper.

---

This PR enhances DSPy’s optimizer suite by enabling adversarial prompt exploration, making it a powerful tool for LLM evaluation and security research. 🚀