Adopt replicate synchronous API for Flux generations

[lzell]

• Added support for Replicate's synchronous API, announced here
• Added deprecation notices to existing Flux-Schnell, Flux-Dev, and Flux-Pro methods that use polling
• Created new methods that the deprecation notices point to. The new methods use http long polling, meaning connections remain open until the prediction result is available. This reduces the number of requests that are made to AIProxy. Results also arrive faster, as we don't have to worry about a coarse polling interval.

[MartinW]

@lzell Thanks for all the work on this, it's much quicker.

replicateService.createPrediction previously allowed any type for the input. I created my own schema to also pass in hfLora, loraScale and guidanceScale.

createFluxSchnellImageURLs does not allow this. Could it? I can’t override it as it's final class.

[lzell]

Whoops, hey @MartinW. I would like to keep generics out of createFluxSchnellImageURLs as it's intended to be easier to use than the the general interfaces. But we can certainly accommodate your needs with a couple approaches:

1. Get rid of all final class definitions so you can customize as you see fit outside of the lib
2. Add the properties that you need to the flux schema. I'm curious, though, because I don't see hf_lora, lora_scale, and guidance_scale on the schnell schema. Are you seeing your image generations affected by those properties?

[lzell]

@MartinW all public classes in the SDK are now open. Thank you for the feedback: https://github.com/lzell/AIProxySwift/pull/65