fix: type api_key as optional on LiteLLMConversation._get_litellm_object

[AmirF194]

Root cause

LiteLLMConversation.get_litellm_object() declared api_key: str, but the method's
own docstring and body document and implement a ValueError for api_key is None.
That branch is real, not hypothetical: set_api_key() forwards its api_key
argument here unchanged, and biochatter/podcast.py calls
set_api_key(api_key=os.getenv("OPENAI_API_KEY"), ...), which is None whenever
that environment variable is unset. The declared type contradicted the implemented
and documented behavior.

Fix

Widened the annotation to api_key: str | None, matching the sibling
implementations in openrouter.py and langchain.py, which already type this
parameter the same way. Also renamed the method to _get_litellm_object, per
@slobentanzer's comment on #306: it is not part of the public API.

Verification

• New tests in test/test_llm_connect/test_litellm_conversation.py: one asserts
  the annotation now includes NoneType (fails on main, passes on this branch),
  one asserts the documented ValueError fires for api_key=None.
• pytest test/test_llm_connect/: 119 passed. 3 pre-existing failures are
  unrelated to this change (the xinference optional dependency group, not
  installed in my local run) and reproduce identically on unmodified main.
• ruff check / ruff format on both changed files: no new violations beyond
  the four SLF001 hits from the rename itself, consistent with how this file
  already tests other private methods directly.
• Not verified: the repo's xinference extra and the full CI matrix, since
  installing torch locally exceeded available disk on my machine.

Fixes #306

[AmirF194]

No rush, just checking in about a week after opening. CI on this fork hasn't run yet since it needs a maintainer to approve the workflow run for a first-time contributor, so let me know if there's anything else you'd like changed in the meantime.

[AmirF194]

Closing to keep the queue clean; happy to reopen if there's interest.