fix partial wrappers and deferred annotations

[technillogue]

#1772 broke partial and partialmethod, as used in https://github.com/replicate/cog-llama-template/blob/main/predict.py#L234-L259. this attempts to fix that while maintaining support for deferred annotations

this PR was ready to be merged except for implementing the test for partialmethod schema generation being an unnecessarily slow integration test instead of a more lightweight test.

[nickstenning]

Does this need to be an integration test? Could it be a worker test instead?

[technillogue]

I was just following #1772, should test_predict_works_with_deferred_annotations be moved as well?

[nickstenning]

Ah, right. I'd be tempted to move both of those into test_worker if it's possible to reproduce the original failure there. It's not really clear to me what the original failure was, though, so it's possible that doesn't work.

[technillogue]

https://replicatehq.slack.com/archives/C05600FDYTE/p1723843667345019

    **get_input_create_model_kwargs(signature, input_types),
      ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/usr/local/lib/python3.11/site-packages/cog/predictor.py", line 306, in get_input_create_model_kwargs
    raise TypeError(f"No input type provided for parameter `{name}`.")
TypeError: No input type provided for parameter `prompt`.

and

  File "/root/.pyenv/versions/3.11.7/lib/python3.11/site-packages/cog/predictor.py", line 374, in get_input_type
    input_types = get_type_hints(predict)
                  ^^^^^^^^^^^^^^^^^^^^^^^
  File "/root/.pyenv/versions/3.11.7/lib/python3.11/typing.py", line 2381, in get_type_hints
    raise TypeError('{!r} is not a module, class, method, '
TypeError: functools.partial(<bound method Predictor.predict of <predict.Predictor object at 0x70abe38f0490>>) is not a module, class, method, or function.

[technillogue]

I think it needs to at least be a predictor test because I don't think worker calls get_input_type, though a schema test would work too

[technillogue]

@nickstenning, if the delay from this test isn't prohibitive, I'd rather just merge this to unblock forks of the old cog-llama-template models, it might make sense to move several of these tests to server/test_predictor.py but that's a fair amount of new stuff

[nickstenning]

Given that you've still got to roll this thing out either way, I'm not sure merging a slow test is going to save much time. Let's not continue to add to the mess. Could we write some better unit tests for get_{input,output}_type instead?

We've leaned heavily into "can't we just" on the cog codebase for quite a while now, and it's got us into a right mess. I'd like us to start doing things we can support for the longer term.