NeuralFloat convolution-backward approximation proofs.

This entry point gathers the bias, input, kernel, and reverse-node bounds used to relate finite-precision convolution gradients to their real-valued specifications.