Fix/duplicate llm call

[Rodrigosilvars83]

Problem

Every chat message triggers two LLM API calls and the frontend receives
the response twice — once as a single token event from the agent loop's
non-streamed call, and again chunk-by-chunk from the trailing stream call.

This causes:

• Visible duplicated text in the UI.
• Doubled token cost on every message.
• Potential message divergence — the second call may produce a slightly
  different answer than the first.

Cause

In chat_completion_stream, when the LLM returns a final answer (no
tool_calls), the loop yields message.content and breaks. But control
then falls through to an unconditional second call
(self.client.chat.completions.create(..., stream=True)) that re-asks
the LLM and yields the same response again.

Fix

Track whether the agent loop produced a final answer (finalized). If it
did, reuse the content that was already yielded and skip the streaming
call. The streaming call now only runs when the loop exhausted
MAX_AGENT_TURNS without finalizing — acting as a fallback to force a
final answer.

Notes

• Also wrapped the trailing stream in try/except so it can't crash silently.
• Small cleanups: consolidated the two if blocks that insert the system
  prompt, continue for unknown tool calls, and for _ in (unused loop var).
• No behavior change for the happy path; only removes the duplication.

Testing

Manually traced the control flow for three scenarios:

1. LLM answers without tool calls → answer yielded once (was twice).
2. LLM uses tools then answers → answer yielded once (was twice).
3. Agent exhausts MAX_AGENT_TURNS → fallback stream produces the final answer.