[trust remote code] Additional features

[sayakpaul]

From the model card of https://huggingface.co/hf-internal-testing/tiny-sdxl-custom-all, if we run

from diffusers import DiffusionPipeline
pipeline = DiffusionPipeline.from_pretrained("hf-internal-testing/tiny-sdxl-custom-all", trust_remote_code=True)

assert pipeline.config.unet == ('diffusers_modules.local.my_unet_model', 'MyUNetModel')
assert pipeline.config.scheduler == ('diffusers_modules.local.my_scheduler', 'MyScheduler')
assert pipeline.__class__.__name__ == "MyPipeline"

pipeline = pipeline.to("cuda")
images = pipeline("test", num_inference_steps=2, output_type="np")[0]

assert images.shape == (1, 64, 64, 3)

it will lead to:

Traceback (most recent call last):
  File "/home/sayak/test.py", line 2, in <module>
    pipeline = DiffusionPipeline.from_pretrained("hf-internal-testing/tiny-sdxl-custom-all", trust_remote_code=True)
  File "/home/sayak/diffusers/src/diffusers/pipelines/pipeline_utils.py", line 1073, in from_pretrained
    cached_folder = cls.download(
  File "/home/sayak/diffusers/src/diffusers/pipelines/pipeline_utils.py", line 1842, in download
    pipeline_class = getattr(diffusers, cls_name, None)
TypeError: getattr(): attribute name must be string

To reproduce, you must clear the cache associated with hf-internal-testing/tiny-sdxl-custom-all.

This PR fixes that.

Additionally, this PR adds the functionality to serialize the module files with the pipeline making it easy for users to share a self-contained pipeline repository on the Hub that can be readily loaded with trust_remote_code=True without having to go through the steps outlined here: #5472 (comment). Here's a gist to test it all out: https://gist.github.com/sayakpaul/685d263df0bd845cbe8997ebd468a5c5.

More comments inline.

[sayakpaul]

Can also parameterize this based on a function argument if that's preferred.

[sayakpaul]

Closing in favor of #5556.