Add the alloc_uninit function. (#1901)

* Add the alloc_uninit function.

* Dummy metal fix.

* Lazy initialization.

candle-core/src/dummy_cuda_backend.rs

@@ -210,6 +210,10 @@ impl crate::backend::BackendDevice for CudaDevice {
         Err(Error::NotCompiledWithCudaSupport)
     }
 
+    unsafe fn alloc_uninit(&self, _shape: &Shape, _dtype: DType) -> Result<Self::Storage> {
+        Err(Error::NotCompiledWithCudaSupport)
+    }
+
     fn storage_from_cpu_storage(&self, _: &CpuStorage) -> Result<Self::Storage> {
         Err(Error::NotCompiledWithCudaSupport)
     }