Fix the reproducibility audit for a training config

An AI said a PyTorch training setup was "fully reproducible" because it called torch.manual_seed(42), yet two runs gave different validation scores. Python's and NumPy's random generators, the CUDA seed, the data loader's shuffling, its worker processes and cuDNN's choice of kernels each need a setting of their own. Your team audits run configs for every one of them.

An assistant wrote missing_determinism_settings(config_lines), the audit. It is in the editor. It has a bug. Find it and fix it.

Input and output:

• config_lines: the run's config, one key=value setting per line. Later lines override earlier ones, as in a layered config file.
• Return the required settings that are missing, in alphabetical order.

Rules the tests check:

• The required settings, in alphabetical order: cudnn_deterministic, dataloader_generator, numpy_seed, python_seed, torch_cuda_seed, torch_seed and worker_init_fn.
• Split a line at its first = into a key and a value. Skip lines with no =. Trim whitespace from both ends of the key and of the value, and lowercase both.
• If a key appears more than once, its last line wins.
• A required setting is missing when its key never appears, or its final value is empty, false or none. Any other value counts, 0 included.
• Other keys are ignored.