[WebNN] Support device tensor for WebGPU EP on ORT backend

Device tensor is already supported for OpenVINO EP on ORT backend in
[1], this CL supports device tensor for WebGPU EP on ORT backend. The
main difference is that device tensor is only used when WebNN developers
do not need to access the underlying data, because the WebGPU EP does
not allow clients to access underlying tensors directly.

[1] https://chromium-review.googlesource.com/c/chromium/src/+/6972837

Bug: 445971854
Change-Id: Iff08fde488117c2c9dabbe7a7bb66d53f850726d
Reviewed-on: https://chromium-review.googlesource.com/c/chromium/src/+/7088044
Reviewed-by: Rafael Cintron <[email protected]>
Reviewed-by: Hu, Ningxin <[email protected]>
Commit-Queue: Wang, Wei4 <[email protected]>
Cr-Commit-Position: refs/heads/main@{#1549642}

Author: wangw-1991

webnn/conformance_tests/tensor.https.any.js

@@ -1330,6 +1330,28 @@ const testDispatchTensor = (testName) => {
         mlContext, outputTensor,
         new Float32Array(sizeOfShape(shape)).fill(1.0));
   }, `${testName} / same constant multiple graphs`);
+
+  promise_test(async () => {
+    // Construct a simple graph: OUTPUT = IDENTITY(INPUT) to test whether the default
+    // tensor is initialized to zero.
+    const builder = new MLGraphBuilder(mlContext);
+    const inputOperand = builder.input('input', {dataType: 'int32', shape: [1024]});
+    const graph = await builder.build({'output': builder.identity(inputOperand)});
+
+    const inputTensor = await mlContext.createTensor({
+      dataType: inputOperand.dataType,
+      shape: inputOperand.shape
+    });
+
+    const outputTensor = await mlContext.createTensor({
+      dataType: inputOperand.dataType,
+      shape: inputOperand.shape,
+      readable: true
+    });
+
+    mlContext.dispatch(graph, {'input': inputTensor}, {'output': outputTensor});
+    await assert_tensor_data_equals(mlContext, outputTensor, new Uint32Array(1024));
+  }, `${testName} / default tensor uninitialized`);
 };
 
 /**