Dataloader

This MR is for data handling. Issue aidge#31 (closed) explain the added features. In aidge_core, the features added are :

1. Stimuli and StimuliImpl - A class wrapping a data sample. Stimuli has two functioning modes. The first mode enables to load data samples from a dataPath and optionnaly store the data in-memory. The second mode enables to store a data sample that was already loaded in memory. StimuliImpl provides the implementation to load a specific file extension.
2. Database - An abstract class representing a map from a key to several Stimulis (for example a stimuli image and a stimuli label). All databases should inherit from this class. All subclasses should overwrite :cpp:function:Database::getItem to fetch data from a given index.
3. DataProvider - Compose batches by fetching data from a given database.
4. Database and DataProvider python biding
5. Connection of tensors to the GraphView handled by Scheduler::connectInputs

ConnectInputs uses the OrderedInputs feature of the graphView.

In the case where data is passed during the call of Scheduler::forward, the call of GraphView::forwardDims (before forward) does not have the input dimensions. Thus GraphView::forwardDims have now an optional argument to pass the dimension of the input tensors.

---

Implement the solution presented in issue aidge_backend_opencv#3 (closed).

• Change the constructor of TensorImpl to take the dimensions of the tensor as argument instead of the number of elements. For tensorImpl and tensorImpl_cpu.
• Adapt pybind_tensor with dimensions instead of number of elements

include/aidge/backend/cpu/data/TensorImpl.hpp

@@ -33,7 +33,7 @@ private:
 public:
     static constexpr const char *Backend = "cpu";
 
-    TensorImpl_cpu(DeviceIdx_t device, NbElts_t length) : TensorImpl(Backend, device, length) {}
+    TensorImpl_cpu(DeviceIdx_t device, std::vector<DimSize_t> dims) : TensorImpl(Backend, device, dims) {}
 
     bool operator==(const TensorImpl &otherImpl) const override final {
         const auto& typedOtherImpl = reinterpret_cast<const TensorImpl_cpu<T> &>(otherImpl);
@@ -47,8 +47,8 @@ public:
         return i == mNbElts;
     }
 
-    static std::shared_ptr<TensorImpl_cpu> create(DeviceIdx_t device, NbElts_t length) {
-        return std::make_shared<TensorImpl_cpu<T>>(device, length);
+    static std::shared_ptr<TensorImpl_cpu> create(DeviceIdx_t device, std::vector<DimSize_t> dims) {
+        return std::make_shared<TensorImpl_cpu<T>>(device, dims);
     }
 
     inline std::size_t scalarSize() const noexcept override final { return sizeof(T); }