Compare analytic gradients with numerical estimates for a scalar computation graph:

Here $a$, $b$, $c$, and $f$ are scalar inputs, $e$ is an intermediate value, and $L$ is the loss. Derive the analytic gradients using the chain rule. Numerical gradients use the supplied positive step h and forward differences, changing one input while keeping the others fixed.

Return a four-item tuple: the loss as a Python float, a list of four analytic Python-float gradients, a list of four numerical Python-float gradients, and their largest absolute pairwise difference as a Python float. Both lists use a, b, c, f order.