feat: add LiteLLM as LLM provider

[RheagalFire]

Summary

• Add LiteLLM as a new LLM provider, giving DemoGPT users access to 100+ LLM providers (Anthropic, AWS Bedrock, Vertex AI, Azure, Groq, Cohere, Mistral, Together, Fireworks, etc.) through the LiteLLM Python SDK
• Works in both the agenthub layer (LiteLLMChatModel for agents/RAG) and the chains layer (via create_llm() factory)
• Activated by passing provider="litellm" to DemoGPT() or using LiteLLMChatModel directly

Changes

• demogpt_agenthub/llms/litellm.py -- new LiteLLMChatModel extending both BaseChatModel (LangChain runnable) and BaseLLM, calls litellm.completion() with drop_params=True
• demogpt_agenthub/llms/__init__.py -- export LiteLLMChatModel (lazy import, no breakage if litellm not installed)
• demogpt/chains/_llm_factory.py -- shared factory that creates LiteLLMChatModel when provider="litellm", ChatOpenAI otherwise
• demogpt/chains/chains.py -- Chains.setLlm() and Chains.getModel() use factory, accept provider param
• demogpt/chains/task_chains.py -- TaskChains.setLlm() uses factory, accepts provider param
• demogpt/chains/task_chains_seperate.py -- same
• demogpt/model.py -- DemoGPT.__init__() accepts provider param, passes through to all chain classes
• pyproject.toml -- add litellm>=1.80.0,<1.87.0 as optional dependency (pip install demogpt[litellm])

Tests

1. Live E2E: LiteLLMChatModel.run() (Azure Foundry -> Claude Sonnet 4.6)

from demogpt_agenthub.llms.litellm import LiteLLMChatModel

model = LiteLLMChatModel(model='azure_ai/claude-sonnet-4-6')
result = model.run('What is 2+2? Reply with just the number.')

Response: 4
Type: str

2. Live E2E: LangChain pipe compatibility (used by agents)

chain = ChatPromptTemplate.from_template('Reply with just: {word}') | model | StrOutputParser()
pipe_result = chain.invoke({'word': 'pong'})

Pipe result: pong

3. Live E2E: Chains factory path

from demogpt.chains._llm_factory import create_llm

llm = create_llm(model='azure_ai/claude-sonnet-4-6', provider='litellm')
result = llm.invoke('Reply with just: pong')

Factory result: pong
Type: LiteLLMChatModel

Example usage

Agenthub (agents, RAG):

from demogpt_agenthub.llms.litellm import LiteLLMChatModel
from demogpt_agenthub.agents import ToolCallingAgent

# Provider API keys read from env: export ANTHROPIC_API_KEY=sk-ant-...
llm = LiteLLMChatModel(model="anthropic/claude-sonnet-4-20250514")
agent = ToolCallingAgent(tools=[...], llm=llm)
result = agent.run("What is the weather in Paris?")

Main DemoGPT app:

from demogpt.model import DemoGPT

# Provider API keys read from env: export ANTHROPIC_API_KEY=sk-ant-...
demo = DemoGPT(
    model_name="anthropic/claude-sonnet-4-20250514",
    provider="litellm",
)

Supported model formats (LiteLLM convention):

• anthropic/claude-sonnet-4-20250514 -- Anthropic
• bedrock/anthropic.claude-sonnet-4-20250514-v1:0 -- AWS Bedrock
• vertex_ai/gemini-2.5-flash -- Google Vertex
• azure_ai/claude-sonnet-4-6 -- Azure AI Foundry
• groq/llama-3.3-70b-versatile -- Groq
• ollama/llama3 -- Local Ollama

Risk / Compatibility

• Additive only -- no existing providers modified
• LiteLLMChatModel follows the exact same dual-inheritance pattern as OpenAIChatModel(ChatOpenAI, BaseLLM), just swapping the LangChain base for a custom BaseChatModel that calls litellm SDK
• provider parameter defaults to None, preserving existing OpenAI behavior
• litellm is an optional dependency -- existing users without it are unaffected (lazy import with try/except)
• drop_params=True ensures unsupported params are silently dropped per provider

[RheagalFire]

cc @melih-unsal