Fail fast on non-retriable LiteLLM status codes

[yangbaechu]

Summary

Fail fast on non-retriable LiteLLM errors instead of retrying them with exponential backoff. This keeps the existing retry behavior for transient failures, while surfacing clear configuration and authorization issues earlier.

Problem

LiteLLMClient.__call_api() currently retries almost all exceptions up to api_max_retry, which means clearly non-retriable failures can spend a long time in exponential backoff before the run fails.

Changes

• treat LiteLLM errors with status codes 401, 403, and 404 as non-retriable
• raise those errors immediately instead of retrying them
• keep the existing retry loop for other errors
• add tests verifying that:
  • retryable errors still retry
  • non-retriable status-code errors fail fast
• skip the LiteLLM-specific tests when the optional litellm dependency is not installed