Run-each-task-in-a-subprocess wrappers repeat the NumPy RNG

A common memory-release pattern runs every training task in a fresh ProcessPoolExecutor(max_workers=1) or Pool(1), forked from a parent that draws nothing itself. Each task then starts from the same legacy np.random global state, while in-process sequential training would give each task a different stream. Unseeded models that shuffle or subsample with np.random (feature permutation, np.random.choice feature subsets, minibatch shuffles) therefore consume one identical stream in every rolling task. Hash np.random.get_state() at the start of each task to detect it. In-process gives N distinct hashes; the wrapper gives 1. Fix: reseed inside the child (np.random.seed(None) or a SeedSequence spawn keyed by task id). Another bug to check in these wrappers: re-wrapping the function inside the task loop nests the subprocesses one level deeper per task.