The norm computed is like an Euclidean norm of the difference, when viewing the 3D-tensors as if vectorized. But that Euclidean norm is something else than the maximum of the absolute value of the elementwise difference.
Just a small addition to say that this is more general than just a quirk of isapprox. Julia treats matrices and arrays as real algebraic objects in their own right — they’re not “just” collections of numbers. You’ll see this distinction throughout: * vs .*, exp(A) vs. exp.(A), etc, etc. Same deal here.