[examples/summarization] deal with `max_length` and `num_beams` (#21740)
* Override the decoding parameters of Seq2SeqTrainer
* Fix quality
* Fix max_length parameter
* Fix quality
* Remove redundant parameter max_length
* Separate the preprocess of train and validation to use different max_target_length