Define "underlying execution device" concept

[anssiko]

The WebNN specification talks about "underlying execution device" in its prose, including in the normative prose, but we don't currently have an agreed upon definition for this concept.

To improve on this, we could add a new <dfn> to refer to and to provide a place where to improve this definition as we gather further implementation experience on the topic.

Related #895

[zolkis]

Lifting some thoughts from #895.

Previously, a context was associated with an "underlying execution device" which used to be one of CPU, GPU, NPU ("cpu", "gpu", "npu" device types).

Nowadays we should rather associate a context with an execution plan that maps specific parts of the model graph to the best available combination of accelerators for inference.

During the build phase, implementations could define an internal, prioritized list of execution providers/delegates based on the context options, which the runtime / underlying platform would use for the actual decisions. Whether or not that list is exposed to apps is a separate concern.

During the dispatch phase, the actual execution plan is finalized for inference. The unsupported ops fall back on the next device, down to a CPU (which is implicit everywhere AFAIK). The proposed graph.devices would expose that list, as discussed in #836.

The text formulation should allow both a single [primary] device per context (with CPU fallback) and supporting heterogeneous sub-graph execution on a set of devices.

Some references:
https://onnxruntime.ai/docs/execution-providers/
https://ai.google.dev/edge/litert/performance/delegates
https://developer.android.com/ndk/guides/neuralnetworks