Trainer - deprecate tokenizer for processing_class (#32385)

* Trainer - deprecate tokenizer for processing_class

* Extend chage across Seq2Seq trainer and docs

* Add tests

* Update to FutureWarning and add deprecation version

examples/pytorch/semantic-segmentation/run_semantic_segmentation.py

@@ -403,7 +403,7 @@ def main():
         train_dataset=dataset["train"] if training_args.do_train else None,
         eval_dataset=dataset["validation"] if training_args.do_eval else None,
         compute_metrics=compute_metrics,
-        tokenizer=image_processor,
+        processing_class=image_processor,
         data_collator=default_data_collator,
     )