Gradient of BCE

Implement:

def bce_gradient(
    input: np.ndarray,
    p: np.ndarray,
    target: np.ndarray,
) -> tuple[np.ndarray, np.ndarray]:
    """
    input:  shape (batch_size, in_features)
    p:      shape (batch_size, 1), sigmoid(linear(input, weight, bias))
    target: shape (batch_size, 1)

    Returns:
        grad_weight: shape (1, in_features)
        grad_bias: shape (1,)
    """

p/target follow 01-hypothesis-function's convention, (batch_size, 1), never squeezed to (batch_size,), so grad_weight/grad_bias come out matching weight's (1, in_features) / bias's (1,) shape directly, the same convention Linear Regression's 03-mse-gradient uses.