fix: guard against empty/filtered LLM responses in openai adapter

[qizwiz]

What

Adds null checks before accessing choices[0].message in three files:

• agentverse/llms/openai.py — generate_response and agenerate_response (4 locations)
• scripts/evaluate_responsegen.py — eval loop
• agentverse/memory/chat_history.py — summarize_memory

Why

The OpenAI SDK can return two silent failure modes that cause unhandled crashes:

1. IndexError — choices is an empty list ([]) when the provider returns no completions
2. AttributeError — choices[0].message is None when content is filtered (e.g. Gemini returns HTTP 200 with PROHIBITED_CONTENT finish reason)

Both crash the agent mid-task with an unhelpful traceback rather than a clear error.

Fix

if not response.choices or response.choices[0].message is None:
    raise ValueError("LLM returned empty or filtered response")

This converts both failure modes into a single, descriptive ValueError that callers can catch and handle.

Detected by pact static analysis.