New operators merge request

New files

Added for the export the kernels, the templates of the config, and kernel files for these operators:

• Sigmoid
• Softmax
• BatchNorm2D
• MatMul
• Reshape
• Gather
• Transpose

Modified files

• All element wise operators now support array broadcasting
• aidge_export_arm_cortexm/_Aidge_Arm/kernels/FullyConnected/aidge_fc_float32.c to better handle different shaped inputs ;
• aidge_export_arm_cortexm/_Aidge_Arm/templates/configuration/fullyconnected.jinja in accordance with the kernel;
• operators.py to include the new operators and change the shape of the input passed to the FC operator, to reflect the changes made to the kernel;

Major modifications

• For the FC operator, the nb_inputs parameter is now computed as nb_channels * channel_height * channel_width with the input always reshaped as [batch_size (usually == 1), nb_channels, feature_height, feature_width]. nb_outputs is now equal to outputs_dims[0][1] with the output_dims staying unchanged. Before, nb_inputs was equal to nb_channels with the input reshaped as [nb_channels, H, W] (essentially removing the batch dimension). This created issues for inputs where the nb_channels = 1 and H,W > 1, only computing the result based on the 1st item of the input array. I believe that this could have been solved only by modifying the way nb_inputs is computed, and carry on discarding the batch size, but this could be useful if someone one day wants to do batched inference.
• Some operations need to know what are the dimensions of the inputs and outputs (array broadcasting, transpose, etc) while others may have arrays as arguments/attributes (permutations indexes for transpose, the indexes for Gather, etc). Those arrays aren't graph nodes, they are parameters of the operator. In order to store and pass them to functions in C, they are saved in a new folder named dimensions where for each array, a file "layer_name_{nature_of_array}.h" is created, with nature_of_array = { DIMS, INDEXES, PERMUTATIONS}

TO-DO

So far, the reshape operator simply copies the data from a pointer to another. This is not optimal.