fix: default EvalHub adapter to stream=false for reliable tool scoring

Non-streaming responses include tool_invocations/tool_calls in the JSON
body. Streaming relies on delta.tool_calls which not all agents emit
(e.g. AutoGen uses a custom mcp.tool_usage SSE event). Defaulting to
false ensures tool scorers work for all agents out of the box; jobs can
still opt in to streaming via job parameters.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

Author: andrewdonheiser

evals/evalhub_adapter/config.py

@@ -96,7 +96,7 @@ class AgenticEvalParams:
     timeout_seconds: float = 30.0
     verify_ssl: bool = True
     fixtures_path: str = "fixtures"
-    stream: bool = True
+    stream: bool = False
 
     # MLflow trace enrichment (reads tool calls from agent-side traces)
     mlflow_tracking_uri: str | None = None