Leaf client implementations can rely on that infrastructure to create canonical AIFunction instances of their own.

I've not seen any need for that. Have you?

We remove a hard dependency on STJ 9 for function factories.

I've also not seen that as a problem for apps, which are the general consumer of this.

This PR proposes we move it to the Abstractions so that:

My main concerns are:

1. These aren't abstractions nor are they needed by the 90% case of implementors. They're typically an app level concern, and apps reference M.E.AI. If you need AIFunctionFactory, there's a super high chance you also need FunctionInvokingChatClient.
2. This is just one creation method. What happens if we decide to add an OpenAPI import method, or for model context protocol, or bringing a prompt file in as a function, or other such imports. Those would naturally belong here as well, but also not in abstractions.
3. We don't currently imbue the generated functions with metrics or other forms of telemetry, but we might want to (SK does). This would either prevent that or force more dependencies into abstractions.
4. At some point I expect we'll add a source generator, and the type as is will get much less use.

Doesn't seem like the risk is worth the reward. And nothing prevents us from type forwarding them down later if we decide then we actually want this.