tests : fix test-thread-safety when compiling with multiple backends

[Acly]

test-thread-safety fails with multiple backends that see the same device (eg. CUDA + Vulkan)

llama_model_load_from_file_impl: skipping device Vulkan0 (NVIDIA GeForce RTX 4070) with id 0000:2b:00.0 - already using device CUDA0 (NVIDIA GeForce RTX 4070) with the same id
llama_model_load_from_file_impl: invalid value for main_gpu: 1 (available devices: 1)
main: failed to load model 'C:\Users\aclys\AppData\Local\llama.cpp\ggml-org_models_tinyllamas_stories15M-q4_0.gguf'

This aligns the test's device enumeration with the one in llama_model_load_from_file to count only unique devices

[slaren]

I think it would be better to set llama_model_params::devices to the device we want to test instead of main_gpu. There is also no reason to skip duplicated devices here, since the goal is to test the backend implementation.

[Acly]

I think it would be better to set llama_model_params::devices to the device we want to test instead of main_gpu. There is also no reason to skip duplicated devices here, since the goal is to test the backend implementation.

Makes sense, I updated the code.