feat: add realtime building blocks to the SDK

[isidentical]

Building realtime apps without these are super hard and makes the code look terrible, this basically abstracts all the required stages for the user and lets them just define the inference behavior similiar to how the regular HTTP stuff works. Real world example of this code clutter: https://github.com/fal-ai/serverless-registry/blob/9f49dad913e10404a647eff3f7fa2f2e79e7f721/registry/image/lcm/inference.py#L292-L431

example:

class RealtimeApp(fal.App, _scheduler="nomad"):
    requirements = ["websockets", "msgpack"]

    def setup(self):
        pass

    @fal.endpoint("/")
    def generate(self, input: Input) -> Output:
        return Output(text=input.prompt)

    @fal.endpoint("/ws", is_websocket=True)
    async def generate_ws(self, websocket: WebSocket) -> None:
        await websocket.accept()
        for _ in range(3):
            await websocket.send_json({"message": "Hello world!"})
        await websocket.close()

    @fal.realtime("/realtime")
    def generate_rt(self, input: Input) -> Output:
        return Output(text=input.prompt)

[badayvedat]

how does the consumer determine whether the response is a msgpack struct or a json? or why don't we use msgpack for all the messages?

[isidentical]

fal emits its own messages as json, so this is done to mimic that. binary protocol is used only for I/O, and errors and other stuff goes as json. the users are expected to decode the frames to see if they are a text frame or a binary frame, and use the right format for that (json and msgpack respectively). https://github.com/fal-ai/fal-js/blob/12887d4b876793ef060924e5bc10a08d39d6313a/libs/client/src/realtime.ts#L473-L495