Add GTests for Conv GPU bias, 1D and 3D paths

[harz05]

Adds GTests exercising three Conv GPU paths that no existing alpaka model covered. All current Conv alpaka models are 2D, group 1, no bias, so these branches in the generated kernels were never run.

• ConvWithBias exercises the BiasBroadcastKernel plus the GEMM beta=1 accumulate. Two output channels with distinct biases (10, 100) so a wrong-channel broadcast is visible.
• Conv1d exercises the fDim==1 branches in the weight-vec and im2col kernels.
• Conv3d exercises the fDim>2 depth branches and the depth handling in im2col.

Each model lists W and bias as initializers with an iota input. Reference outputs are verified against PyTorch (F.conv1d / F.conv2d / F.conv3d). Tested on Google Colab (NVIDIA T4, CUDA 12.x): all Conv-dim tests pass with no regressions.

Steps to reproduce on Colab

!apt-get install -y libprotobuf-dev protobuf-compiler libgtest-dev
!git clone https://github.com/harz05/SOFIE.git && cd SOFIE && git checkout feat/conv-dim-coverage-tests
!cd SOFIE && mkdir build && cd build && cmake -Dtesting=ON -DCMAKE_INSTALL_PREFIX=../install -DCMAKE_BUILD_TYPE=RelWithDebInfo -DENABLE_ALPAKA_TESTS=ON -DALPAKA_BACKEND=cuda .. && cmake --build . --target install -j$(nproc)
!cd SOFIE/build && ctest -V

Test models

ConvWithBias : input x[1,1,5,5] = iota, weight W[2,1,3,3] all ones, bias [10, 100], kernel [3,3], pads [1,1,1,1], output y[1,2,5,5]. The two distinct biases make a wrong-channel broadcast visible.

Conv1d : input x[1,2,7] = iota, weight W[3,2,3] = iota, bias [1, 2, 3], kernel [3], pads [1,1], output y[1,3,7].

Conv3d : input x[1,1,3,4,4] = iota, weight W[2,1,2,3,3] = iota, bias [5, 50], kernel [2,3,3], pads [0,1,1,0,1,1] (no depth pad, h/w padded by 1), output y[1,2,2,4,4].

W and bias are ONNX initializers (baked into the .dat), so infer() takes only x. Iota weights give each output and input channel a distinct slice so a wrong index is visible. References computed with PyTorch (cross-correlation, no kernel flip) reading W and bias straight from the committed .onnx.