Qualia-codegen-core

New features

- Add support for separate scale factor for bias.
- Add support for 1D/2D grouped convolutions (incl. depthwise convolutions).
- Add left shift overflow detection when `TRAPV_SHIFT` is defined and not using CMSIS-NN.
- TorchModelGraph: add support for ReLU6 activation layer.
- TorchModelGraph: add support for `operator.add()` function in addition to Add layer.
- TorchModelGraph: add support for `torch.nn.functional.adaptive_avg_pool2d` function.
- TorchModelGraph: add support for `torch.flatten()` function in addition to Flatten layer.
- Add NucleoL452REP example project using CMake instead of STM32CubeIDE.
- Add common libqualia-neuralnetwork for string to float decoding, quantization of inputs and classification task results, used only for NucleoL452REP CMake project for now.
- Abort compilation in case of unsupported activation function for layer.

Bug fixes

- TorchModelGraph: fix AdaptiveAvgPool1d layer.
- Linux example: fix single.cpp build.
- SparkFunEdge example: fix round-to-nearest support.

New features

- Add support for nearest rounding mode for weights and activations in all supported layers and for inputs in all included example targets, with and without CMSIS-NN/NMSIS-NN.
- Generate a new `include/defines.h` file to be used as a global pre-processor macros definition file (used to set global rounding mode for CMSIS-NN/NMSIS-NN for now).
- Add an `exclude` argument to `Quantizer.quantize_weights()` to exclude some weights type by name in a custom plugin.

Bug fixes

- Longan Nano: build NMSIS-NN from upstream sources instead of using pre-built binary in order to apply custom optimizations and support floor rounding mode instead of nearest.
- `graph.layers`: fix weights property inheritance

Initial release of Qualia-CodeGen-Core.