[Fix] Export results differ to much from the backend cpu

Context

This Issue points out a great difference between the backend cpu results and the export cpp ones which has to be reduced.

This MR fixes the differences observed in quantized exports, but there still is a difference of results for fp32 exports.

Model	fp32	int8
LeNet	✔️	✔️
Resnet18	❌	✔️

On the table above, validation is done using the aidge_cmp option, which evaluate the results between the aidge_backend_cpu and aidge_export_cpp kernels, with an absolute precision of 10^-5 (for float exports).

Changes

Fix

• The input tensor was not properly transposed, leading to incoherent results;
• The AddAct metaoperator was always inheriting from the QElemWise one. It was then always using the int32_t accumulation type, leading to wrong results in fp32 export;
• The regex recipes were not correctly defined for the Conv DW based meta operators;
• Biases default type was set to float32_t instead of float;
• The convolution kernel was not working properly with a dilation of 2;
• The CI was broken due to a PyTorch update.

Improvement

• The aidge_cmp headers are not compiled anymore if the AIDGE_CMP flag is not set to true, reducing the compilation time.