fix: fix tool-calls finish reason conversion in streaming mode for newer gemini models like gemini-3.5-flash

[hustxiayang]

Description

Problem

When the model returns a tool call, OpenAI expects the finish reason to be tool_calls while Gemini returns STOP. We convert it via

if len(toolCalls) > 0 {
			return openai.ChatCompletionChoicesFinishReasonToolCalls
		}

This was correct for streaming mode because the finish reason and the tool calls are in the same chunk in the previous Gemini models. However, for newer Gemini models like gemini-3.5-flash, they are split into two chunks. The functionCall chunk has
no finishReason, and a trailing chunk carries finishReason: STOP with an empty
text part and no functionCall:

// raw GCP chunk 1 — has the tool call, no finishReason
{"candidates":[{"content":{"parts":[{"functionCall":{"name":"get_current_weather","args":{"...":"..."}},
                                     "thoughtSignature":"..."}]}}],
 "usageMetadata":{"trafficType":"ON_DEMAND"}}

// raw GCP chunk 2 — terminal STOP, empty text, no functionCall
{"candidates":[{"content":{"parts":[{"text":""}]},"finishReason":"STOP"}],
 "usageMetadata":{"promptTokenCount":107,"candidatesTokenCount":31,"thoughtsTokenCount":148,"...":"..."}}

In this case, the conversion would be invalid. Thus, we need to remember whether the model returned a tool call during the streaming mode.

[codecov-commenter]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 84.85%. Comparing base (0553786) to head (fc38b85).

Additional details and impacted files

@@            Coverage Diff             @@
##             main    #2399      +/-   ##
==========================================
+ Coverage   84.84%   84.85%   +0.01%     
==========================================
  Files         151      151              
  Lines       22086    22090       +4     
==========================================
+ Hits        18739    18745       +6     
+ Misses       2215     2214       -1     
+ Partials     1132     1131       -1     

☔ View full report in Codecov by Harness.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.
• 📦 JS Bundle Analysis: Save yourself from yourself by tracking and limiting bundle sizes in JS merges.

[aabchoo]

/retest