WebNN: Implement MLBuffer dispatch

Adds support to execute MLGraphs using MLBuffers. Allows the WebNN
developer to directly bind MLBuffers as input/outputs to graphs
for execution, which keeps MLBuffer data on-device after execution
completes. In future CLs, dispatch can be further optimized.

* Moves out per graph resources required by Dispatch().
* MLGraphBuilder.build no longer pre-allocates I/O.

webmachinelearning/webnn#482

Bug: 1472888
Change-Id: I7400704cf60c149c47c20f22c50d5f12bff89cf9
Cq-Include-Trybots: luci.chromium.try:win11-blink-rel

webnn/conformance_tests/buffer.https.any.js

@@ -14,6 +14,7 @@ if (navigator.ml) {
   testDestroyWebNNBuffer('destroyTwice');
   testReadWebNNBuffer('read');
   testWriteWebNNBuffer('write');
+  testDispatchWebNNBuffer('dispatch');
 } else {
   test(() => assert_implements(navigator.ml, 'missing navigator.ml'));
 }