Cifar models: remove explicit parameters initialization

except for the case of VGG, our parameter initialization code was
matched the default pytorch initialization (per torch.nn operation),
so writing the initialization code ourselves can only lead to
more code and maintenance; and also we would not benefit from
improvements that occur at the pytorch level (e.g. if FB finds a
better initialization for nn.conv2d than today's kaiming init, we
would not benefit).
The VGG initialization we had was "suspicious" and so reverting
to the default seems reasonable.