ML Inference on heterogeneous architectures using SOFIE

SOFIE generates standalone C++ inference code from trained neural networks, letting HEP frameworks run ML models without heavy external dependencies. Its integration with the alpaka library enables portable GPU inference that works across NVIDIA, AMD and CPU backends from a single codebase. The current GPU path covers basic operations but lacks support for several key operators needed to run full neural networks on GPU. This project implements the missing GPU kernels and extends convolution to handle batched inputs, so complete CNN architectures can run end-to-end on GPU through SOFIE. The final goal is a ResNet-style model running fully on the alpaka GPU backend and validated against the CPU reference.