fix: retry transient streamable-http MCP failures on isolated session

[elainegan-openai]

Summary

• retry a streamable-HTTP MCP tool call on a fresh isolated session when the shared session hits a transient transport failure
• keep shared-session healthy-path parallelism; only isolate on the failure path
• add focused regression tests for cancelled sibling calls, 5xx retry, 4xx no-retry, and preserved outer cancellation

Why

The likely remaining bug is earlier in the chain than MCPUtil.invoke_mcp_tool(): one shared-session streamable-HTTP request failure can still fan out into sibling request cancellations. This patch keeps the normal shared session on the healthy path, but retries just the affected call on an isolated session when the shared session sees a transient transport failure.

Concretely, the isolated retry triggers for:

• asyncio.CancelledError
• httpx.ConnectError
• httpx.TimeoutException
• httpx.HTTPStatusError with status >= 500
• BaseExceptionGroup containing one of the above

It does not retry 4xx errors.

Validation

• uv run pytest -q tests/mcp/test_client_session_retries.py
• uv run mypy src/agents/mcp/server.py tests/mcp/test_client_session_retries.py
• ruff check src/agents/mcp/server.py tests/mcp/test_client_session_retries.py

[seratch]

Closing this in favor of #2703