Use He weight initialization instead of Xavier

[tail-call]

See ctgnnlib/training.py:

def init_weights_xavier(m: nn.Module):
    "Usage: `model.apply(init_weights_xavier)`"
    # XXX For ReLU (as opposed to tanh and sigmoid)
    # XXX He initialization is more appropriate
    # XXX
    # XXX <https://arxiv.org/abs/1502.01852>
    # XXX     Delving Deep into Rectifiers: Surpassing
    # XXX     Human-Level Performance on ImageNet Classification
    # XXX
    # XXX     Kaiming He, Xiangyu Zhang, Shaoqing Ren, Jian Sun

    if isinstance(m, nn.Linear):
        init.xavier_uniform_(m.weight)
        m.bias.data.fill_(0.01)

So a He initialization should be used instead.