Partition sample indices from zero through $N-1$ into $k$ folds for cross-validation. When $N$ is not divisible by $k$, the first $N\bmod k$ validation folds receive one extra index.

For fold $i$, validation indices are $F_i$ and training indices are all remaining folds. If shuffle is true, shuffle indices with np.random.default_rng(seed).permutation before partitioning. Return a list of $k$ dictionaries, each containing train_idx and val_idx as one-dimensional integer NumPy arrays.