Dilated Causal Convolution

WaveNet-style models use causal, dilated convolutions: output $t$ only sees inputs at $t, t-d, t-2d, \dots$

y_t = \sum_{k=0}^{K-1} w_k\, x_{t - k d}

with $x_{<0} = 0$ (left padding), so the output has the same length as the input. Implement causal_dilated_conv1d(x, w, dilation) for 1-D arrays.