get_pipeline function passes Path object rather than PretrainedTokenizer

When using certain pipeline tasks (e.g., "image-to-text") the get_pipeline function passes in `model_dir` as the tokenizer ink wargs. However, the Pipeline class expects a [PretrainedTokenizer](https://github.com/huggingface/transformers/blob/bffac926ca6bc6c965a92bfbfd00c567a2c0fb90/src/transformers/pipelines/base.py#L763) object instead. Because `model_dir` is being passed, when the [tokenizer is decoded](https://github.com/huggingface/transformers/blob/bffac926ca6bc6c965a92bfbfd00c567a2c0fb90/src/transformers/pipelines/image_to_text.py#L176) in the subclass it gives the error `"message": "str object has no attribute decode"`.

This should be changed so that this function is using the proper tokenizer for the task.

Start at the get_pipeline entry point and trace how model_dir is passed into tokenizer kwargs. Compare the expected tokenizer in transformers' pipelines/base.py with decoding in pipelines/image_to_text.py, then verify that image-to-text uses a proper tokenizer and no longer raises the reported decode error.

Written by the indexing model from the issue text.