[modular] fallback to default_blocks_name when loading base block classes in ModularPipeline

[yiyixuxu]

so that we are able to correctly load repos like https://huggingface.co/diffusers/flux2-modular

When creating a pipeline using SequentialPipelineBlocks.from_blocks_dict() and saving it, the modular_model_index.json stores _blocks_class_name as "SequentialPipelineBlocks". On loading, from_pretrained tries to instantiate SequentialPipelineBlocks() directly, which has empty block_classes and produces a broken pipeline.

from diffusers import ModularPipeline
pipe = ModularPipeline.from_pretrained("diffusers/flux2-modular")
print(pipe.blocks)

under main, it won't be able to map to Flux2AutoBlocks

>>> print(pipe.blocks)
SequentialPipelineBlocks(
  Class: ModularPipelineBlocks

  Description: 


  Sub-Blocks:
)

This PR added a fallback behavior in ModularPipeline.__init__ method so if the resolved blocks class is not importable from our library or has empty block_classes (i.e., it's a base class like SequentialPipelineBlocks, ConditionalPipelineBlocks, etc.), fall back to the pipeline's default_blocks_name .

[HuggingFaceDocBuilderDev]

The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.

[sayakpaul]

Was just thinking about this today. Thanks!

[sayakpaul]

This is perfect test! I will try to make it available to other pipelines in #13192.