make KolorsPipelineFastTests::test_inference_batch_single_identical pass on XPU

[faaany]

What does this PR do?

Since there is no difference can be seen from the images generated with single and batched inputs, I adjusted the expected maximum difference from 5e-4 to 5e-3 to make the test pass on both XPU and CUDA.

pls help review it. Thanks! @hlky

[hlky]

Thanks @faaany. Same comment as in #11308 (review) but I will add that having expected_max_diff per backend is certainly not a priority and the tolerances mentioned in that comment are not the best example of when this should be a priority. Essentially we wouldn't want to increase expected_max_diff to the point where changes affecting precision may not be detectable, however there haven't been many of these changes required and only for specific tests, test_inference_batch_single_identical in this case, so it would likely be picked up in another test, and we generally confirm outputs are visually the same when making any changes to a pipeline/model anyway.