Feat/add litellm provider

[prodmanpd]

Summary

Add LiteLLM as a provider with a native SDK code path, giving users direct access to 100+ LLM providers (Anthropic, Google, Azure, Bedrock, Groq, Ollama, and more) for document translation without needing an OpenAI-compatible proxy.

Changes

• docutranslate/agents/agent.py - Added _call_litellm_sync() and _call_litellm_async() methods that use litellm.completion() / litellm.acompletion() instead of raw httpx POST. All 4 HTTP call sites patched (send, send_async, _continue_fetch, _continue_fetch_async) to route through litellm when provider == "litellm"
• docutranslate/agents/provider/provider.py - Added "litellm" to ProviderType
• docutranslate/agents/thinking/thinking_factory.py - Added thinking mode delegation by model name for litellm
• tests/test_litellm_provider.py - 8 unit tests

Implementation details

When --provider litellm is set, the Agent bypasses raw httpx and calls litellm's SDK directly:

def _call_litellm_sync(self, data: dict) -> dict:
    import litellm
    kwargs = {
        "model": data["model"],
        "messages": data["messages"],
        "temperature": data.get("temperature"),
        "top_p": data.get("top_p"),
        "stream": False,
        "drop_params": True,
    }
    if self.key and self.key != "xx":
        kwargs["api_key"] = self.key
    if data.get("response_format"):
        kwargs["response_format"] = data["response_format"]
    kwargs = {k: v for k, v in kwargs.items() if v is not None}
    response = litellm.completion(**kwargs)
    return response.model_dump()

• drop_params=True silently handles cross-provider parameter incompatibilities (e.g. frequency_penalty rejected by Anthropic, response_format rejected by some providers)
• litellm.acompletion() used for async path so concurrent translation batches work
• API key forwarded when set, otherwise litellm reads provider env vars (ANTHROPIC_API_KEY, OPENAI_API_KEY, etc.)
• Thinking mode resolved by model name since litellm proxies multiple providers (e.g. qwen-plus gets Qwen's enable_thinking, gemini-2.5-flash gets Google's reasoning_effort)
• Response is model_dump()'d to dict so existing token counting and response parsing works unchanged

Tests

8 unit tests covering provider detection, thinking mode delegation, and type system.

Full regression: 64 passed (56 existing + 8 new), 0 failures.

Example usage

# Direct SDK usage - litellm routes to the right provider
pip install litellm
export ANTHROPIC_API_KEY=sk-ant-...

docutranslate translate \
  --base-url http://localhost:4000/v1 \
  --model anthropic/claude-sonnet-4-6 \
  --provider litellm \
  input.pdf

from docutranslate.agents.agent import Agent, AgentConfig

agent = Agent(AgentConfig(
    base_url="unused-for-litellm",
    model_id="anthropic/claude-sonnet-4-6",
    provider="litellm",
))
# Agent uses litellm.completion() instead of raw httpx
# No OpenAI-compatible proxy needed

Risk / Compatibility

• Existing providers untouched - litellm path only activates when provider == "litellm"
• Raw httpx path unchanged for all other providers
• litellm is imported lazily inside the method bodies