tolerance as 1e-2 is too loose, @kaixuanliu, pls do more investigations.

This PR is aimed to solve two precision failed test cases in diffusers CI on Intel XPU platform:
1.pytest -rA tests/pipelines/kandinsky2_2/test_kandinsky_controlnet.py::KandinskyV22ControlnetPipelineIntegrationTests::test_kandinsky_controlnet
2.pytest -rA tests/pipelines/kandinsky3/test_kandinsky3_img2img.py::Kandinsky3Img2ImgPipelineFastTests::test_encode_prompt_works_in_isolation

for case 1, I compared the output images, they are perceptualy same(see attached output.bmp
expected_output.bmp
), so I think the tolerance value change is accepatable.

for case 2, maybe the 1e-2 value is loose, I will take it back. Let's solve case 1 first.

For case2, pls refer to the analysis and fix in #12474

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