Improve Pad to match ONNX and overcome current limitations

Refactor Pad operator to solve several issues.

Should fix:

• aidge_onnx#72 (closed);
• aidge_onnx#81 (closed);
• aidge_backend_cuda#53 (closed): backward support in CUDA;
• aidge_backend_cuda#19 (closed): Pad1D support in CUDA;
• aidge_backend_cpu#46 (closed) and previous associated MR !447 (closed);
  • Implemented solution: add a specific ConstantLowest pad mode, instead of trying to work around the correct type and value of the constant_value input Producer. It can be handled very easily in exports or backends, much like the Constant pad mode.
• #244 (closed) and aidge_onnx#25 (closed): count_include_pad support. Also close aidge#290 (closed);
  • Implemented solution: introduce a ignorePads attribute to AvgPooling op, like suggested here. It is extremely simple to implement and allows to fully uncouple padding and pooling when count_include_pad is true.
• aidge#131 (closed): Pad1D support in CUDA;
• Missing backward implementation on CPU;
• Correct un-reported bug in Reflect mode.