Why use: tanh(a + b) VS tanh(a) + tanh(b) and why use a single layered neural network layer

Question

In this machine translation exercise:

First question:

why does this class BahdanauAttention (see class at end) do th