[tests] Generalize alpaka tests over backends and implement the hip build

[harz05]

The alpaka tests were hardcoded for the CUDA backend even though the generated code and the CMake ALPAKA_BACKEND option are already backend-agnostic. This PR makes the test sources compile for any alpaka backend and also wires the hip build.

Changes

• test/alpaka/TestAlpakaCommon.h: the accelerator tag (TestTag) is now picked from the ALPAKA_ACC_GPU_*_ENABLED macro the build defines, and the fixture types (TestAcc, TestDev, TestQueue) derive from it via alpaka::TagToAcc. Removed the cuda_runtime.h include and an unused nvml.h include.
• Test bodies: Session<alpaka::TagGpuCudaRt> becomes Session<TestTag>, and cudaDeviceSynchronize() becomes alpaka::wait(device). The device-wide wait is kept rather than removed because the generated Session owns its own queue, so waiting on the test queue alone is not enough.
• test/CMakeLists.txt: the source list and target definition are shared across backends, with only the toolchain, flags and libraries selected per backend. ALPAKA_BACKEND=hip now builds a TestCustomModelsFromONNXForAlpakaHip target (LANGUAGE HIP, links hipblaslt hipblas amdhip64) instead of failing with "not implemented". The cuda build is unchanged: same target name, flags and libraries as before.
• Fixed the backend validation, which compared the input against itself, so any lowercase value passed and valid uppercase spellings were rejected. It now checks against the list of known backends.

Test names are unchanged, so existing CI and logs stay comparable. The cpu and sycl stubs are left as they were.

Testing

• CUDA on H100
• HIP on MI100