Improve export

Miscellaneous fixes and changes to robustify the exports, which will fix aidge_export_cpp#32 (closed) and aidge_onnx#64 (closed).

• Fixed I/O order handling in scheduler_export();
• Handle data format: remove hugly transpose fix in Producer export;
• Handle data format: remove format trick in FC kernel;
  • Support both Default/NCHW and NHWC format in two FC kernels;
  • Added a recipe to automatically transpose the weights during export if possible.
• Make data format a "weak" property of tensors:
  • By default, data format change is ignored if the tensor dimension does not match the format;
  • When tensors are resized, if the tensor new dimension does not match the format, it is reset to Default.
• Improved display of GraphView::save() to display type and format even when the tensor is undefined.
• Handle optional outputs:
  • Allow to remove graph outputs with GraphView::setOrderedOutputs();
  • Take it into account during export;
  • Make adaptToBackend() comply with that;
  • Loaded ONNX graph outputs are conformant to ONNX outputs (no extra outputs are added anymore);
  • Recipes ExpandMetaOp(s) and AdaptToBackend do not introduce new outputs anymore.
• Handle multi-precision:
  • Memory manager now handles bytes instead of words;
  • Memory is allocated in bytes;
  • Added support for memory alignment, with a default memory alignment of 16 bytes (adapted up to 64-bits systems);
  • Export memory array is now always of type unsigned char.
• Enforce loaded ONNX model data format to NCHW;
• Added NHWC format support in forwardDims() for Pad operator;
• Add result values checks in lenet.py;
• AdaptToBackend should not rename existing nodes (@axelfarr);
• Fixed createUniqueName() should never return an empty name;
• Fixed #279 (closed) with an assert.