[train][Test] Clean up http server setup in unit test

[CharlieFRuan]

Remove redundant code by extracting out set_up_http_server() in skyrl-train/tests/gpu/gpu_ci/test_inference_engine_client_http_endpoint.py

[gemini-code-assist]

Code Review

This pull request refactors the HTTP server setup logic in the unit tests by extracting a set_up_http_server() helper function. This is a good improvement that reduces code duplication and improves maintainability.

My review includes two main points:

1. A potential NameError in the finally blocks of the refactored tests. The server thread and port variables are used for cleanup but might not be initialized if an error occurs earlier in the try block. I've suggested a pattern to make the cleanup more robust.
2. An improvement to the port finding logic within the new helper function to use a more idiomatic approach.

[gemini-code-assist]

The variables server_port and server_thread are defined inside the try block. If an exception occurs before their assignment (e.g., during init_inference_engines), they will not be defined, and this finally block will raise a NameError, preventing proper cleanup. To make the cleanup robust, you should initialize these variables to None before the try block and check for their existence here.

This same issue exists in other refactored tests in this file: test_http_endpoint_completions_routing_and_batching, test_http_endpoint_openai_api_with_weight_sync, test_http_endpoint_with_remote_servers, and test_http_endpoint_error_handling.

        if server_port is not None:
            shutdown_server(host=SERVER_HOST, port=server_port, max_wait_seconds=5)
        if server_thread is not None and server_thread.is_alive():
            server_thread.join(timeout=5)

[gemini-code-assist]

This implementation for finding an available port can be simplified and made more idiomatic by binding to port 0, which lets the operating system choose an available ephemeral port. This avoids manually iterating through a range of ports.

Note that both the current implementation and the suggested one have a time-of-check-to-time-of-use (TOCTOU) race condition, as the port is released after being found and before the server binds to it. A more robust solution would involve passing the bound socket to the server, but that would require changes to the serve function and is likely out of scope for this PR.

    def _find_available_port(host: str) -> int:
        """Find an available port by binding to port 0."""
        import socket
        with socket.socket(socket.AF_INET, socket.SOCK_STREAM) as s:
            s.bind((host, 0))
            return s.getsockname()[1]

    # Find an available port
    server_port = _find_available_port(SERVER_HOST)