relu -> linear -> relu -> linear, added back to the input. The second
linear layer starts near zero so the block starts as the identity, the same
trick made_module() uses on its output heads and the reason a deep stack
trains at all from a cold start.