Assuming that pads=[0, 0, 0, 0] in case it is missing from the ONNX model.

Add tests which simultaneously test onnx frontend and 2pc/3pc/cpp backend.

Previosuly we were iterating through the entire program for each identifier's
name patching. This was a quadratic algorithm resulting in >30mins compile time
for deep networks. Now we just do a linear scan of the program once and patch
expressions which need patching. Compile time is reduced to seconds.

Resolving the [ABY] std::bad_alloc() error which was being encountered on running dot_product example

Resolving the [ABY] std::bad_alloc() error which was being encountered on running dot_product example

If distro doesn't provide >3.13, we manually build cmake

Now we can compile models similar to how we compile tensorflow models. The
server can compile the model with:
  ~/EzPC/Athos/CompileONNXGraph.py --config config.json --role server

This generates a zip file which contains:
  optimised_model.onnx
  config.json

The optimised_model.onnx is the same model but is a pruned version and doesnt
contain model weights. This zip file can then be sent to the client and client
can compile with:
  ~/EzPC/Athos/CompileONNXGraph.py --config config.json --role client