Make it easier to write good docs for model inputs

Have you ever come across a Replicate model input like this one?

🤔 What is "classifier-free guidance"? Surely some experts know, but not me. I'll go look in the README. Hmm it's not in the README. I guess I just won't change that setting! Since I don't know what it will do, It's not worth it to change the value, wait for a prediction, then try find out what effect it had.

If the description looked more like this, I'd be more likely to experiment with changing its value:

Putting more detail in the input description itself means that users see it in lots of useful contexts, like when they're:

• filling out the web form
• using "Intellisense" in their $EDITOR
• referring to a model's API docs

On the other hand, as a model author, the current Python DSL(?) for documenting an input makes writing a long description feel a bit awkward:

def predict(self,
    guidance_scale: float = Input(
        description="I don't want to make this too long because if I do the line will go on forever and push the other inputs out of view and this just feels unholy to write a single-line string this long I guess I could put it in a variable but that would be overkill maybe?", ge=1, le=50, default=7.5
    ),
)

I'm opening this issue to brainstorm on how we can make it easier for model authors to "do the right thing" (write good docs) when creating the interface to the model's (Cog) interface.

Start by reviewing the linked predict.py example and the Python Input DSL shown in the issue, then consider how descriptions appear in the README, web form, editor assistance, and API documentation. Done should mean an agreed, documented approach that makes writing detailed model-input descriptions easier for model authors.