Allow Union and List input types

[mattt]

Alternative to #968 #1292

This PR extends Cog to support List and Union type inputs.

For example, given the following predictor:

from cog import BasePredictor, Path

from typing import List, Union

class Predictor(BasePredictor):
    def predict(self, args: Union[int, List[int]]) -> int:
        if isinstance(args, int):
            return args
        else:
            return sum(args)

The following assertions are satisfied:

resp = client.post("/predictions", json={"input": {"args": 123}})
assert resp.status_code == 200
assert resp.json()["output"] == 123

resp = client.post("/predictions", json={"input": {"args": [1, 2, 3]}})
assert resp.status_code == 200
assert resp.json()["output"] == 6

resp = client.post("/predictions", json={"input": {"args": "abc"}})
assert resp.status_code == 422

resp = client.post("/predictions", json={"input": {"args": ["a", "b", "c"]}})
assert resp.status_code == 422

[bfirsh]

Marvellous. Could do with some docs, but we can fix up later.

I wonder how this interacts with Replicate?

[mattt]

@bfirsh Union[str, List[str]] is expressed in OpenAPI as {"anyOf": [{"type": "string"}, {"type": "array", "items": "string"}}]. Currently, Replicate renders inputs of that schema as a single text field. We can eventually support multiple inputs, but this seems sufficient for now, since users can send multiple inputs via the API.

[peter65374]

only List[int] is supported so far? Does input type support List[Path] or List[str]? I am expecting to input multiple images to process as below

def predict(
        self,
        target_image: Path = Input(
            description="Faces in target image would be changed.",
        ),
        model_faces: list[Path] = Input(
            description="Model Faces would be swapped into target image.", 
            default=None,
        ),
        inference_mode: str = Input(
            default="swap",
            choices=["swap", "detect"],
            description="Face swap mode or detection mode. Default is swap.",
        ),
    ) -> Union[List[tuple], List[Path]]:
        """Run a single prediction on the model"""
        # Check if inference_mode is detect or swap
        if inference_mode == "detect":
            detect_mode = True
        else: 
            detect_mode = False