128. Detect Gradient Explosion

Given a list of per-layer gradient norms (ordered from input layer to output), return the index of the first layer whose norm is more than threshold times the running maximum of all previous layers (and a baseline of 1.0).

Signature: def find_exploding_layer(grad_norms: list, threshold: float = 100.0) -> int

For index i, the running max baseline is max(grad_norms[:i] + [1.0]). If grad_norms[i] / running_max > threshold, return i. Return -1 if no layer explodes.