Need to understand how WebNN supports implementation that involves multiple devices and timelines

[wchao1115]

During the conversation on WebNN and WebGPU interop session, at the WebGPU F2F meeting in San Francisco, Myles Maxfield made an observation about the fact that WebNN MLContext is a single-device context. He raised the awareness that Apple's CoreML can distribute a workload within a single ML graph across multiple devices including the ANE, and said that it would be nice if WebNN has a provision to support that situation as well.

I think this feedback is related to a discussion in the working group of late about ways to create a notion of a default MLContext (when there is no explicit device associated with the context -- either a CPU device or a WebGPU device) that is more open to the implementation. This default context could effectively be whatever the implementer wants it to be including a multi-device context in the case of the Apple's WebKit implementation over CoreML.