enable t5 in SD3 pipe testing and flux img2img and inpaint

[eaidova]

What does this PR do?

• enable text_encoder_3 in SD3 testing for resolving Follow-up request after #916 due to issues on the OpenVINO side #965
• enable flux img2img pipeline (was introduced in diffusers 0.31.0, previous version was based on 0.30.0)
• added test that save_pretrained returns correct model_index.json

Before submitting

• This PR fixes a typo or improves the docs (you can dismiss the other checks if that's the case).
• Did you make sure to update the documentation with your changes?
• Did you write any new necessary tests?

[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.

[nikita-savelyevv]

After I run

optimum-cli export openvino -m katuni4ka/tiny-random-flux ./tmp

I don't see a model for text_encoder_3. Based on PR description I would expect that it should be there. Am I getting something wrong?

[nikita-savelyevv]

Why is flux addition conditioned on transformers version and not on is_diffusers_version(">=", "0.31.0") as above?

[eaidova]

problem in optimum-intel gone not from diffusers version, in github actions there is division by transformers version, diffusers always used the latest. problem with models imports comes from saving text encoders that are part of transformers with upper version of tokenizers than supported in transformers < 4.40 as the result test will fail in precommit

[eaidova]

@nikita-savelyevv flux does not have text_encoder_3, it is expected. there is no changes in model export part for text_encoder_3, only its ignoring removed in tests. During writing tests I found that compilation of small t5 model used in test sd3 model failed with segfault on compilation stage, but this issue does not appear in real big models, so I disable loading of this model in tests i(it is optional part of sd3 pipeline and model can work without it, just with small accuracy drop). Now in 2024.5 issue was fixed

[nikita-savelyevv]

Please take a look at failed diffusion tests