Add OpenResponses API client implementation

[simonw]

This adds llm/responses.py with sync and async model classes that
implement the OpenResponses API using httpx as the transport layer:

• ResponsesModel (sync) and AsyncResponsesModel (async) extending
  KeyModel/AsyncKeyModel for integration with the llm framework
• Pydantic models for ResponseResource and streaming events
• SSE streaming parser for real-time text deltas
• Custom error classes (ResponsesAPIError, ResponsesAuthenticationError,
  ResponsesRateLimitError, ResponsesInvalidRequestError)
• Support for tool/function calls
• Comprehensive test suite (35 tests) using TDD approach

[simonw]

I manually tested it with this:

import llm
from llm.responses import ResponsesModel
from llm.default_plugins.openai_models import Chat
from llm.tools import llm_time


key = llm.get_key("openai")

chat_model = Chat(
    "gpt-5-mini",
    vision=True,
    reasoning=True,
    supports_schema=True,
    supports_tools=True,
)
print("=== chat model ===")
print(chat_model.chain("what is the time?", key=key, tools=[llm_time]).text())

print("=== responses model ===")

responses_model = ResponsesModel(
    "gpt-5-mini", "https://api.openai.com/v1", supports_schema=True, supports_tools=True
)
print(responses_model.chain("what is the time?", key=key, tools=[llm_time]).text())

It didn't work at first because tool results were not passed correctly, but I fixed that.