fix(cuda): address review feedback on NVIDIA Blackwell/PRIME detection

[Copilot]

Applies four fixes from the code review on the NVIDIA Blackwell/PRIME detection changes, covering log correctness, spurious warnings, incorrect device indexing, and unsafe env-var overrides.

Changes

• model_manager.cpp — single LOG call per GPU entry: The NVIDIA GPU startup log was built across 3–4 chained LOG(...) calls, risking interleaved output under concurrent access. Suffix string is now assembled first and emitted in one call.

• system_info.cpp — conditional fallback warning: The nvidia-smi detection failed warning was emitted unconditionally, even on machines with no NVIDIA driver. Moved inside the fs::exists("/proc/driver/nvidia/gpus") check so it only fires when the kernel module is actually loaded and the fallback is genuinely attempted.

• system_info.cpp — drop synthetic gpu.index in /proc fallback: GPUInfo::index documents the physical nvidia-smi device index. The /proc path was assigning a filesystem-iteration ordinal, which could cause wrong device selection. index is now left at -1; uuid is the identifier for this path.

• llamacpp_server.cpp — respect existing __NV_PRIME_RENDER_OFFLOAD_PROVIDER: The provider was unconditionally hard-coded to NVIDIA-G0 whenever __NV_PRIME_RENDER_OFFLOAD was unset, silently overriding any user-set value. Now gated on the provider also being unset:

const char* existing_provider = std::getenv("__NV_PRIME_RENDER_OFFLOAD_PROVIDER");
if (!existing_provider || existing_provider[0] == '\0') {
    env_vars.push_back({"__NV_PRIME_RENDER_OFFLOAD_PROVIDER", "NVIDIA-G0"});
}

[kenvandine]

wrong target branch