fix: forward base_url to instructor client for OpenAI-compatible endpoints

[shivam2407]

Summary

Fixes #5204

When using an OpenAI-compatible provider with a custom base_url (e.g. self-hosted vLLM, Ollama, or any non-OpenAI endpoint), InternalInstructor silently discards the base_url and sends structured output requests to api.openai.com instead of the configured endpoint.

Root cause

_create_instructor_client() extracts only model and provider from the LLM, then calls instructor.from_provider(f"{provider}/{model_string}"). The base_url is never forwarded, and instructor.from_provider() creates a default client pointing at api.openai.com/v1/.

Fix

Added _get_llm_client_kwargs() to extract base_url, api_base, and api_key from the LLM object. When base_url is present, constructs an explicit SDK client instead of using from_provider():

Provider	Client used	Key forwarded as
openai	OpenAI(base_url=...)	base_url
azure / azure_openai	AzureOpenAI(azure_endpoint=..., api_version=...)	azure_endpoint
anthropic	Anthropic(base_url=...)	base_url
Others	from_provider() fallback	api_key only

Additional behaviors:

• api_base is normalized to base_url; when both are set, base_url takes precedence
• All guards use is not None (not truthiness) so empty strings are handled correctly
• When no custom base_url is set, behavior is unchanged (from_provider() path)

Changes

File	Change
internal_instructor.py	Added _get_llm_client_kwargs(), _create_instructor_client_with_base_url() with per-provider client construction. Updated _create_instructor_client() to route through the new path when base_url is present.
test_converter.py	8 new tests + fixed 1 existing mock

Test plan

• 8 new tests pass (OpenAI/Anthropic/Azure forwarding, api_base normalization, collision precedence, string LLM, no-base_url default)
• All 50 existing tests still pass
• Manual testing with vLLM/Ollama endpoint + output_pydantic task