use valid padding in first conv layer of synthesis generator

[bodokaiser]

If we use padding='same' for the first convolution layer with kernel_size=9 we get output shape [batch_size, 16, 16, 16, 1] as intended from the original paper.