Fix HIP classification in backend device test

[Blakeolson21]

What changed

backend_device_resolution_test mirrors the backend registry mapping from src/framework/core/backend.cpp, but it still treated the ROCm registry as CUDA after HIP became its own BackendType.

On a HIP build, the test would discover ROCm, request a CUDA backend, and fail instead of validating the enumerated HIP device.

This change:

• maps ROCm to BackendType::Hip in the test contract
• adds the missing HIP label
• checks that an unregistered HIP backend is rejected, alongside CUDA, Vulkan, and Metal

Validation

Both focused builds and test runs passed on an Apple M3 Max with AppleClang 17.0.0:

macOS CPU core build
checked 0 accelerator registries
backend_device_resolution_test passed

macOS Metal core build
checked 1 accelerator registry
backend_device_resolution_test passed

The prior project warning about the unhandled BackendType::Hip enum value is gone. The remaining compile warning is in the vendored ggml CPU backend.

I did not have a usable HIP host for this change. A Windows CUDA attempt reached commit 4584379101c765a64fac1abc4d1e8f72fc42d2f4, but the host lacked the official CUDA Toolkit, so I am not claiming Windows GPU coverage.

[0xShug0]

Merged!