Handle textual responses when tool_calls is empty

[IloDan]

Problem

Some OpenAI-compatible APIs return message.tool_calls as an empty list
for final textual responses.

The previous parser only fell back to message.content when iterating
over tool_calls raised an exception. With tool_calls=[], it instead
produced model_responses=[] and discarded the valid textual response.

Fix

• Normalize message.tool_calls with message.tool_calls or []
• Preserve message.content when no tool calls are present
• Stop the function-call loop when a final textual response is received

Validation

Tested with Qwen3-8B through a vLLM OpenAI-compatible endpoint on BFCL v4
memory subsets.

Before the fix, all memory subset scores were 0.

After the fix:

• memory_kv: 4.52%
• memory_rec_sum: 29.03%
• memory_vector: 9.68%
• average memory accuracy: 14.41%

This is close to the 14.62% reported for Qwen3-8B FC on the BFCL leaderboard.