ai/core: support reason + "array"

[bernaferrari]

Feature Description

Sequel to #1486.

I wish I could have a reason field before the array, so that if I had

z.object({
reason: z.string(),
arr: z.array(...)
})

The array would stream in chunks and work with array but still allow other fields that are not array. So, ideally,

• if z.array => stream in chunks (current behavior)
• else => stream naturally? Not sure.

Could also be like "if something is inside an array, wait for it to finish to stream, else stream naturally"

[lgrammel]

@bernaferrari I don't want to hardcode something like reasoning, but I'd love to figure out a more flexible way to support array streaming. One thought that I had is to use a validator schema for intermediate results (partial result would only be sent if they pass that schema), but this will not solve the "is-object-completed" issue. The "is-object-completed" detection is tricky, because it needs to be specified to which level it applies to. Lmk if you have ideas.