feat: add LiteLLM as AI gateway provider

[RheagalFire]

Pull Request type

• Feature

./gradlew spotlessApply has been run.

Changes in this PR

Adds LiteLLM as a new AI gateway provider, giving Conductor users access to 100+ LLM providers through a single unified OpenAI-compatible interface.

LiteLLM is a user-deployed AI gateway/proxy that unifies 100+ providers (OpenAI, Anthropic, Azure, Bedrock, Vertex AI, Mistral, Cohere, etc.) behind one OpenAI-compatible endpoint. Users deploy it locally, via Docker, Kubernetes, or on AWS/GCP/Azure using Terraform modules, and point Conductor at it. This lets users route requests through one config, add cost tracking/rate limiting/fallbacks, and switch providers without changing workflow definitions.

Files added:

• ai/src/main/java/.../providers/litellm/LiteLLM.java - new provider implementing AIModel via OpenAICompatChatModel (same pattern as Grok/Perplexity). Zero new dependencies.
• ai/src/main/java/.../providers/litellm/LiteLLMConfiguration.java - Spring @ConfigurationProperties(prefix = "conductor.ai.litellm"). base-url is required (LiteLLM is user-deployed, no public SaaS endpoint). api-key is optional (many deployments run without auth).
• ai/README.md - added LiteLLM to supported providers table (now 13), configuration docs, environment variables, and drop_params guidance.

Configuration:

conductor.ai.litellm.base-url=http://litellm-proxy:4000
conductor.ai.litellm.api-key=${LITELLM_API_KEY}  # optional, only if proxy has auth enabled

Full E2E (Conductor server in Docker + Postgres + LiteLLM proxy -> Azure Foundry / Anthropic Claude Sonnet 4.6):

POST /api/workflow/execute/litellm_e2e_test/1

Workflow: COMPLETED
Task: LLM_CHAT_COMPLETE -> COMPLETED
  Result: 4
  Tokens: 25
  Finish: STOP

Complete chain verified: Conductor workflow -> LLM_CHAT_COMPLETE task -> LiteLLM provider -> OpenAICompatChatModel -> HTTP to LiteLLM proxy -> Azure Foundry -> Anthropic Claude -> response parsed back through the full Conductor pipeline.

Example workflow:

{
  "llmProvider": "litellm",
  "model": "gpt-4o",
  "messages": [{"role": "user", "message": "What is the capital of France?"}],
  "temperature": 0.7,
  "maxTokens": 500
}

Set drop_params: true in your LiteLLM proxy config (litellm_settings) so provider-unsupported parameters (e.g. frequency_penalty for Anthropic) are silently dropped instead of causing 400 errors.

[RheagalFire]

cc @v1r3n @manan164