Need support for Deconvolution (Transposed convolution, ONNX ConvTranspose) node

[vuzelac-cadence]

Want to discuss the approach for supporting Deconvolution, needed for upsampling in some segmentation networks.

I plan to add support for ONNX ConvTranspose which nicely maps onto our backend Deconvolution. But not sure what that node should map to in CPU/Interpreter

One approach is to reuse ConvolutionGrad node (and it's libjit implementation) and just skip unrelated things (like bias gradient) by adding a flags. However, I am not sure about the performance of this implementation - Convolution is supported in various libraries and leaving this one with only the default implementation seems slow.

Alternatively, input can be prepared to perform regular convolution (slicing and scattering of input, transposing weights and what not), which again is not optimal just like the above.

We don't care much about any efficient implementation in CPU/Interpreter (except for the simulation speed during development) but wanted to hear if you have any plans here.