Add knowledge distillation (KD) loss to aidge_learning

Context

Adding the Knowledge Distillation (KD) loss to aidge for incremental learning.

Modified files

• aidge_learning/src/loss/distillation/KD.cpp: definition of the KD loss (highly inspired by BCE loss)
• aidge_learning/include/aidge/loss/LossList.hpp: Adding the KD loss the loss list
• aidge_learning/python_binding/learning/loss/pybind_Loss.cpp: Add the biding python for the KD Loss

TODO

Integrate the merge request into the main branch

• NOT DONE
• DONE
• TO DO

include/aidge/loss/LossList.hpp

@@ -34,6 +34,17 @@ Tensor MSE(std::shared_ptr<Tensor>& prediction,
 
 Tensor BCE(std::shared_ptr<Tensor>& prediction,
            const std::shared_ptr<Tensor>& target);
+/**
+ * @brief Compute the Knowledge Distillation loss.
+ * This function returns the loss and set the ``grad()`` of the prediction
+ * input.
+ * @param student_prediction Tensor returned by the Aidge Graph of student model,
+ * it is important that this tensor is not a copy as otherwise the backward 
+ * function will not have a gradient to start.
+ * @param teacher_prediction Tensor returned by the Aidge Graph of teacher model.
+ */
+Tensor KD(std::shared_ptr<Tensor>& student_prediction,
+          const std::shared_ptr<Tensor>& teacher_prediction, float temperature = 2.0f);
 
 Tensor CELoss(std::shared_ptr<Tensor>& prediction,
            const std::shared_ptr<Tensor>& target);