Change default kernel_initializer for convolutional layers.

[GhassenJ]

Currently, all convolutional under edward2/tensorflow/layers take 'trainable_normal' as the default kernel_initializer.
A change to 'trainable_deterministic' or a less sensitive initializer than 'trainable_normal' should be applied to all conv layers.

[dustinvtran]

We typically use trainable_he_normal and with the stddev_initializer set to a very small value. That could help.