Fix flaky test_a2a_error_handling test

The test was using a fixed 0.1-second sleep and assuming the task would
fail within that time. In slower CI environments or under Python 3.10,
this timing assumption could fail.

Changed to use the same retry loop pattern used by all other a2a tests:
- Poll the task status up to 50 times (5 seconds total)
- Break early when the task reaches 'failed' state
- Raise clear error if timeout is reached

This matches the pattern established in previous commits (a253fad, etc.)
that fixed similar flakiness in other a2a tests.

🤖 Generated with [Claude Code](https://claude.com/claude-code)

Co-Authored-By: Claude <[email protected]>

Author: conradlee

tests/test_a2a.py

@@ -515,8 +515,14 @@ def raise_error(_: list[ModelMessage], info: AgentInfo) -> ModelResponse:
             task_id = result['id']
 
             # Wait for task to fail
-            await anyio.sleep(0.1)
-            task = await a2a_client.get_task(task_id)
+            max_attempts = 50  # 5 seconds total
+            for _ in range(max_attempts):
+                task = await a2a_client.get_task(task_id)
+                if 'result' in task and task['result']['status']['state'] == 'failed':
+                    break
+                await anyio.sleep(0.1)
+            else:  # pragma: no cover
+                raise AssertionError(f'Task did not fail within {max_attempts * 0.1} seconds')
 
             assert 'result' in task
             assert task['result']['status']['state'] == 'failed'