as we weren't supporting padding, we neglected to use them in the plain
evaluation, but this result in a confusing error message for the user
when trying to use padding. This fixes the problem by using padding
properly during evaluation, and leaves the error up to the compiler.
ONNX spec:
"kernel_shape : list of ints
The shape of the convolution kernel. If not present, should be
inferred from input W."
We were taking the number of input/output feature maps into account,
which we realized we should not.
- 1D, 2D, and 3D convolutions
- conversion to MLIR is supported only for 2D convolution
- not full support yet: still limitations on group and other params
