Fixed forkserver still skips __main__ preload under python -m and Jupyter

CPython 3.13.8+/3.14.1+ forkserver preloads __main__ only when spawn.get_preparation_data() emits 'init_main_from_path', i.e. when the program was started from a script file. For python -m pkg it emits 'init_main_from_name', which forkserver.main ignores. So the server never imports the main module, and workers get independent np.random state even when the main module imports numpy.random/pandas. The same holds when __main__ has no __file__: python -c, the REPL, and IPython/Jupyter kernels. I checked with a real ipykernel driven by jupyter_client: 16/16 distinct draws on 3.14.1, while the same code as a script gave duplicates. So when you assess the impact of worker-RNG duplication on 3.14, ask how the user launches the code. Notebook users of a library are mostly unaffected; batch scripts run by path are affected. Console-script entry points are files, so they are preloaded.