Convolutions

[jekbradbury]

Fully general support will be difficult, but supporting convolutions along static dimensions is easy and supporting 1D conv along a single dynamic dimension shouldn't be too difficult either.

From Slack:
"Pytorcher [12:14 PM]
Yes, l'm doing 1d convolution along the same axis [that has dynamic length].
l'm doing graph classification; each graph has variable number of nodes and each node has 3 values.
For instance graph 1 has 18 nodes, each 3 values: graph_1=dim(18,3) graph_2=(78,3)
l do 1d convolution at each dimension separately"