[refactor] SD3 docs & remove additional code

[a-r-r-o-w]

Refactors SD3:

• Removes F.scaled_dot_product_attention check which is already done in the processor (see here)
• Accesses parameters directly instead of via self.config for init methods
• Updates docs
• Updates type hints from FloatTensor -> Tensor since the former is deprecated

[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.

[yiyixuxu]

thanks!

[yiyixuxu]

is this on purpose?

[a-r-r-o-w]

Indeed, yes! All the latest integrations don't document the internal blocks. This is because of two reasons:

• the parameters are already documented for main transformer, so copying for each internal block is redundant and almost always has introduced mismatched explanations (from doing the refactoring, i sometimes see parameters that don't exist are documented, and sometimes the explanations are just wrong)
• The internal blocks are not user-facing API or linked in main diffusers documentation, so they will not show up there anyway