Link against ggml in bin so we can get the available devices without loading a model.

gpt4all-backend/llmodel.h

@@ -101,6 +101,7 @@ public:
     virtual bool initializeGPUDevice(int /*device*/) { return false; }
     virtual bool hasGPUDevice() { return false; }
     virtual bool usingGPUDevice() { return false; }
+    static std::vector<GPUDevice> availableGPUDevices();
 
 protected:
     // These are pure virtual because subclasses need to implement as the default implementation of