Remove zero initialization of to_attn_bias weights, since for these bias=False

[amorehead]

Line 8 of Algorithm 24 uses a LinearNoBias layer following LayerNorm to merge pairwise with attention bias representations, as follows

However, the code has the LinearNoBias weights also initialized with zeros, meaning both the weights and biases of these modules are initially zero (or null) which leads to these weights receiving no gradients throughout training.