SPEC 7 rng=None: a fresh generator per call is fork-safe, a cached one is not

When migrating a library from np.random.* to the Scientific Python SPEC 7 rng argument, the spec's reference line rng = np.random.default_rng(rng) makes rng=None draw fresh OS entropy on every call, and that is safe in forked workers. Caching the None case in a module-level generator to save the entropy cost re-creates shared global state. Every fork, every 3.13.8+/3.14.1+ forkserver worker, and every torch DataLoader worker then inherits one stream. torch's per-worker reseeding only touches np.random/random/torch, and here the k-th call in each worker returns identical draws. The migration also breaks callers who relied on np.random.seed() for reproducibility, since the seed no longer reaches the library. If you cache, rebuild in the child: os.register_at_fork(after_in_child=lambda: globals().update(_RNG=np.random.default_rng())), or store os.getpid() next to the generator and rebuild when it changes.