Fix failing DeepSpeed model zoo tests

[pacman100]

What does this PR do?

1. Fixing DeepSpeed model zoo tests
   a. tests/deepspeed/test_model_zoo.py::TestDeepSpeedModelZoo::test_zero_to_fp32_zero3_img_clas_vit was failing earlier due to ValueError: --label_column_name label not found in dataset 'hf-internal-testing/cats_vs_dogs_sample'. Make sure to set --label_column_name to the correct text column - one of image, labels.. Fixed this.
   b. tests/deepspeed/test_model_zoo.py::TestDeepSpeedModelZoo::test_zero_to_fp32_zero3_trans_fsmt was failing earlier due to TypeError: Old language model head is of type <class 'torch.nn.modules.sparse.Embedding'>, which is not an instance of <class 'torch.nn.modules.linear.Linear'>. You should either use a different resize function or make sure that old_lm_head are an instance of <class 'torch.nn.modules.linear.Linear'>.. This was because the resize_token_embeddings didn't account for the case when lm_head is an instance of torch.nn.Embedding and it was forcing the new lm_head to be an instance of torch.nn.Linear.
   c. tests/deepspeed/test_model_zoo.py::TestDeepSpeedModelZoo::test_zero_to_fp32_zero2_trans_m2m_100 was failing earlier due to ValueError: --max_source_lengthis set to 1024, but the model only has 512 position encodings. Consider either reducing--max_source_length to 512 or using a model with larger position embeddings. Fixed the test to use appropriate source and target sequence lengths.

[muellerzr]

Thanks for fixing! cc @amyeroberts for final review :)

[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.

[ArthurZucker]

Thanks, tests are unrelated and I think failed on main.
Would be nice to add a test, but alright otherwise

[pacman100]

Would be nice to add a test, but alright otherwise

Could you please clarify what you mean?

[amyeroberts]

Thanks for fixing all of these! 🙏