fix(bfcl): use message.content when tool_calls is empty (OpenAI-APIs)

[pcabriada]

Summary

_parse_query_response_FC in OpenAICompletionsHandler used a try/except around iterating message.tool_calls. When tool_calls is None, iteration fails and the code correctly falls back to message.content. When some OpenAI-compatible servers (e.g. vLLM) return tool_calls: [] for text-only assistant turns, the list comprehension succeeds and yields model_responses == [], so content is never read. That breaks multi-step FC flows (e.g. agentic / memory traces collapsing to empty steps).

Change

• Read tool_calls once.
• If truthy, parse tool calls and IDs as before.
• Otherwise set model_responses from message.content and tool_call_ids to [].

Motivation / context

• OpenAI-style APIs often use tool_calls: null when there are no tool calls; some gateways use an empty list instead.
• Behavior is unchanged when there is at least one tool call.

How to test

From berkeley-function-call-leaderboard after pip install -e ., with a configured FC model:

bfcl generate --model <MODEL_NAME> --test-category simple_python --num-threads 1