Extract and preserve provider-specific headers (llm_provider-*) when
handling error responses from LLM providers. This ensures that useful
debugging information from providers is available even when requests
fail with BadRequestError or similar exceptions.