Describe the bug

Hi,

There's a bug in pipeline_utils.py which causes pipeline.from_pretrained to fail if the pipeline was partially downloaded. Specifically the code doesn't handle missing components in the feature_extractor, safety_checker, scheduler and tokenizer folders, at least on Windows platform.

The cause of the bug is this line

allow_patterns += [os.path.join(k, "*") for k in folder_names if k not in model_folder_names]

This turns some folder names into a regexp pattern but on Windows the path joining is done via {parent}\\{child}, which gives a pattern like this

[
'text_encoder/model.safetensors', 
'vae/diffusion_pytorch_model.bin', 
'vae/diffusion_pytorch_model.safetensors', 
'text_encoder/pytorch_model.bin', 
'unet/diffusion_pytorch_model.safetensors', 
'unet/diffusion_pytorch_model.bin', 
'feature_extractor\\*', 
'safety_checker\\*', 
'scheduler\\*', 
'tokenizer\\*'
]

The \\* pattern doesn't play nice with the regexp matching later and causes some files to be incorrectly excluded from the "consider list", after the expected_files = [f for f in expected_files if any(p.match(f) for p in re_allow_pattern)] call, expected_files is only

[
'unet/diffusion_pytorch_model.bin', 
'vae/diffusion_pytorch_model.bin', 
'text_encoder/pytorch_model.bin', 
'model_index.json'
]

And this means if some files are missing in those folders mentioned above, diffusers will not even try to download them and causes loading errors down the road.

To reproduce:

• Use Windows
• Call StableDiffusionPipeline.from_pretrained("stabilityai/stable-diffusion-2-1-base") and wait until the pipeline is loaded
• Delete the feature_extractor folder from the pipeline cache folder C:\Users\<YOU>\.cache\huggingface\hub\models--stabilityai--stable-diffusion-2-1-base\snapshots\<HASH>
• Call StableDiffusionPipeline.from_pretrained("stabilityai/stable-diffusion-2-1-base") again and observe the error

Fix:
I've tried changing the line above to below and it seems to fix the bug for me. This should be safe for non-Windows platforms as well as that's how path joining works for them in the first place.

allow_patterns += [f"{k}/*" for k in folder_names if k not in model_folder_names]

Reproduction

See above

Logs

No response

System Info

diffusers 0.16.1