Expose built-in models

[mmccool]

Platforms (browsers, operating systems, local model servers) will also provide models that it would be useful to access from web applications. How can we do this in a way consistent with downloadable models, so that if the model exists on the platform, the download can be avoided?

One suggestion: Make such models "appear" in the graph (full model) cache and usable from the WebNN API as "pre-built" models.

Related issue: webmachinelearning/webnn#1