feat: add LiteLLM as AI gateway provider

[RheagalFire]

Summary

Adds LiteLLM as an AI provider for embedding, LLM, and vision capabilities. Users select litellm as their provider and use provider-prefixed model IDs (e.g. anthropic/claude-sonnet-4-6) to access 100+ LLM providers through a unified Python SDK. drop_params=True is always passed so provider-unsupported kwargs are silently dropped.

Follows the existing provider pattern exactly: LiteLLMEmbedding, LiteLLMChat, and LiteLLMVision extend the abstract base classes, registered in the registry dispatch functions.

Changes

• app/ai/providers/litellm_provider.py - LiteLLMEmbedding, LiteLLMChat, LiteLLMVision (all three capabilities)
• app/ai/providers/base.py - added LITELLM = "litellm" to ProviderType enum
• app/ai/registry.py - registered litellm in _get_embedding_class, _get_llm_class, _get_vision_class + added label
• pyproject.toml - added litellm>=1.80.0,<2.0 under [project.optional-dependencies].litellm

Verification

Live E2E against Azure Foundry (Claude via LiteLLM SDK):

$ python3 -c "
config = ProviderConfig(
    provider=ProviderType.LITELLM,
    api_key=os.environ['ANTHROPIC_FOUNDRY_API_KEY'],
    model_id='anthropic/claude-sonnet-4-6',
    base_url=os.environ['ANTHROPIC_FOUNDRY_BASE_URL'],
)
llm = LiteLLMChat(config)
result = await llm.generate('Say OK and nothing else.', temperature=0)
"

content: 'OK'
test_connection: ok=True, msg=OK — model=anthropic/claude-sonnet-4-6, response='OK'

Lint clean:

$ ruff check app/ai/providers/litellm_provider.py app/ai/providers/base.py app/ai/registry.py
All checks passed!

Usage

from app.ai.providers.base import ProviderConfig, ProviderType
from app.ai.providers.litellm_provider import LiteLLMChat

config = ProviderConfig(
    provider=ProviderType.LITELLM,
    api_key="sk-ant-...",                    # or leave empty for env-var auth
    model_id="anthropic/claude-sonnet-4-6",  # any litellm model ID
    base_url=None,                           # optional: litellm proxy URL
)

llm = LiteLLMChat(config)
result = await llm.generate("Summarize this document...")

# Tool calling works too:
turn = await llm.generate_with_tools(messages, tools)

Notes

• Additive only. No existing providers modified.
• litellm is an optional extra - lazy-imported inside function bodies.
• generate_with_tools is implemented with full tool-call parsing, matching the OpenAI provider's AssistantTurn output format.
• Vision provider includes retry logic (3 attempts with backoff), matching the OpenAI vision pattern.
• No catalog entries added - the maintainer can add specific litellm model specs as needed.

[RheagalFire]

cc @nduckmink

[nduckmink]

Thank you for your PR, I will review it soon!