fix CI bug for kandinsky3_img2img case

[kaixuanliu]

For full pipeline, related cut_context is set to True in L511 while for pipe_without_text_encoders, this param is set to False by default L116, hence there will be some minor difference between their outputs(in UNet L492 the mean value w/ and w/o cut_context is different). For platforms such as Intel XPU, they use their seperate generator L145, hence the default tolerance value 1e-4 is not enough. Here we explicitly set encode_prompt_inputs["_cut_context"] = True to align with the behavior of full pipeline. @DN6 , pls help review, Thx!

[kaixuanliu]

@a-r-r-o-w@DN6 ， Hi, can you help review? Thx!

[regisss]

@kaixuanliu Does this only affect XPU? Or GPU too?
Can you also provide the test command to reproduce it please?

[kaixuanliu]

@regisss This PR also affects GPU, although on GPU, the tolerance value is OK. Steps to reproduce:
pytest -rA tests/pipelines/kandinsky3/test_kandinsky3_img2img.py::Kandinsky3Img2ImgPipelineFastTests::test_encode_prompt_works_in_isolation

[DN6]

@kaixuanliu better solution here is to change the default value to True in the encode_prompt signature since it's hardcoded in the pipeline anyway.

[kaixuanliu]

@DN6 I agree. Have updated the code.

[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.

[DN6]

Thanks 👍🏽