[Feature Request] Add "output_hidden_states" option to convert_generation.py

### Describe the feature request

I can successfuly use the [convert_generation.py](https://github.com/microsoft/onnxruntime/blob/0c6037b5abe571fc43a55ef7a9d2f846820fbe5d/onnxruntime/python/tools/transformers/convert_generation.py) script to create a T5 code generation model with beam search (my base model is [Salesforce/codet5-base-multi-sum](https://huggingface.co/Salesforce/codet5-base-multi-sum)). However, I'd also like to output the `encoder_hidden_states` output from the encoder subgraph to use in a downstream task. [Huggingface](https://huggingface.co/docs/transformers/internal/generation_utils#transformers.generation_utils.BeamSearchDecoderOnlyOutput) has an `output_hidden_states` flag to append this tensor to the output, but I don't see this option here. I noticed that the `output_sequence_scores` and `output_token_scores` options exist like they do in Hugginface, so I was wondering if there are any plans to implement this option. I've tried manually manipulating the graph to no avail.

### Describe scenario use case

Usage of `encoder_hidden_states` in downstream tasks.

Start with onnxruntime/python/tools/transformers/convert_generation.py and trace the existing output_sequence_scores and output_token_scores options. Compare those paths with the requested Hugging Face output_hidden_states behavior and determine how encoder_hidden_states should be exposed for beam-search generation. Done means the conversion script accepts the option and generated outputs include the encoder hidden states for downstream use.

Written by the indexing model from the issue text.