fix: skip OpenAI model registration when no API key is configured

[pintaste]

Problem

When no OpenAI API key is set, llm still registers ~50 OpenAI models (and several embedding models) at startup. This causes several issues:

• llm models output is cluttered with models the user cannot use
• Model IDs like gpt-4o-mini collide with identically-named models from other plugins (e.g. a LiteLLM proxy)
• The default model resolves to an OpenAI model, so bare llm invocations fail with No key found for users who don't use OpenAI at all

Fix

Add a guard at the top of register_models and register_embedding_models in the built-in OpenAI plugin. If no key is available (neither in keys.json nor OPENAI_API_KEY env var), registration is skipped entirely.

if not llm.get_key(alias="openai", env="OPENAI_API_KEY"):
    return

Users who later configure a key will see the models appear automatically on the next invocation.

Tests

• Added test_openai_models_not_registered_without_key — verifies OpenAI Chat models are absent from llm models when no key is set
• Added test_openai_models_registered_with_key — verifies OpenAI Chat models appear when a key is present
• Updated conftest.py to set a fake OPENAI_API_KEY in the global env_setup fixture so existing tests that exercise OpenAI model features continue to pass
• Added a file-level autouse fixture in test_cli_openai_models.py for explicitness

Fixes #1445