metal : support virtual devices (#18919)

* metal : support virtual devices

* cont : manage buffer type context memory

* metal : add events

* cont : implement cpy_tensor_async

src/llama-context.cpp

@@ -317,6 +317,7 @@ llama_context::llama_context(
                 auto dev_type = ggml_backend_dev_type(ggml_backend_get_device(backend.get()));
                 if (dev_type == GGML_BACKEND_DEVICE_TYPE_CPU) {
                     // ignore CPU backend
+                    // TODO: should we ignore ACCEL types too?
                     continue;
                 }
                 auto * dev = ggml_backend_get_device(backend.get());