One Gradient-Descent Update

Implement:

def gd_step(
    weight: np.ndarray,
    bias: np.ndarray | None,
    grad_weight: np.ndarray,
    grad_bias: np.ndarray | None,
    lr: float,
) -> tuple[np.ndarray, np.ndarray | None]:
    """
    Apply one gradient-descent update.

    Returns:
        updated_weight, updated_bias
    """

Your function should:

1. Move weight and bias a step in the direction that reduces loss, scaled by lr, using the update rule from Theory.
2. Return new values rather than mutating the input arrays in place.
3. If bias is None (no bias parameter, matching 01-hypothesis-function and 03-mse-gradient), return None for updated_bias too, there is nothing to step.