fix(openai): emit incremental tool-call events during streaming

[xisongtang]

Summary

The OpenAI-compatible provider only emitted a tool_call event once the stream finished (when finish_reason was tool_calls/stop). The Google, Anthropic, and Ollama providers all surface tool-call progress as it arrives during streaming, so OpenAI was the inconsistent outlier.

This emits the accumulated tool calls on every chunk that carries tool-call deltas, so consumers can render live tool-argument progress (e.g. streaming UI previews of structured tool output) instead of waiting for the entire response to finish.

Before / After

 // Accumulate streamed tool calls
 if (delta.tool_calls) {
     for (const tc of delta.tool_calls) { /* accumulate into toolCallAccum */ }
+    // Incremental tool-call progress: emit accumulated arguments as they
+    // stream in, matching the Google/Anthropic/Ollama providers.
+    if (toolCallAccum.size > 0) {
+        const calls = Array.from(toolCallAccum.values())
+            .map((tc) => this.normalizeToolCall(tc));
+        yield { type: 'tool_call', calls };
+    }
     if (loopGuard.detection) break;
 }

The existing finish-gated emit (Emit tool calls when stream finishes) is unchanged, so consumers relying on a single final event are unaffected.

Why

Discovered while building live streaming previews: the OpenAI provider gave no intermediate tool_call events, so partial tool arguments (e.g. a large JSON payload being generated) could not be previewed until completion — unlike the other three providers.

Notes

• dist/ is gitignored; only the source change is included.
• I couldn't run bun test in my environment (bun not installed locally). Please run the suite; happy to add a unit test (fetch/SSE mock) if you'd like coverage for this path.