Fix non-deterministic T5 outputs in HiDream pipeline tests

[kaixuanliu]

This PR tries to fix 2 failed test cases for hiream pipeline tests:

tests/pipelines/hidream_image/test_pipeline_hidream.py::HiDreamImagePipelineFastTests::test_cpu_offload_forward_pass_twice
tests/pipelines/hidream_image/test_pipeline_hidream.py::HiDreamImagePipelineFastTests::test_sequential_offload_forward_pass_twice

Root cause: In get_dummy_components(), T5EncoderModel(config) creates the model in training mode by default. The tiny-random-t5 config has dropout_rate=0.1, so each forward pass produces different embeddings due to active dropout — even under torch.no_grad(). In production this isn't an issue because from_pretrained() automatically calls .eval(), but we need to add this explicitly for text_encoder_3 in the test file

[kaixuanliu]

@DN6 @sayakpaul pls help review, thx!

[DN6]

Thanks @kaixuanliu!

[kaixuanliu]

The failed check are pre-existing and unrelated to this change.