feat: Fix LiteLLM provider detection and improve chat completion step naming

[viniciusdsmello]

Summary

• Add LITELLM_PREFIX_TO_PROVIDER_MAP to correctly resolve the provider when models are accessed via a LiteLLM proxy (e.g. gemini/gemini-2.5-flash now maps to Google instead of OpenAI)
• Always use provider-based naming for chat completion steps (e.g. "Google Chat Completion") instead of accepting the generic run_name from the caller (e.g. "Language Model")

Test plan

• Test with LiteLLM proxy serving Gemini models — verify provider shows Google instead of OpenAI
• Test with LiteLLM proxy serving Anthropic models — verify provider shows Anthropic
• Test direct OpenAI usage — verify provider still shows OpenAI
• Verify chat completion step names use provider prefix (e.g. "Google Chat Completion")

🤖 Generated with Claude Code