Batched longform generation from Huggingface Transformers

[noe]

Recently, a couple of improvements were added to the Whisper implementation in the Huggingface Transformers library for batched long form generation. Here is a Twitter thread describing them. The PRs are:

• [Whisper] Add sequential longform decoding huggingface/transformers#27492
• [Whisper] Finalize batched SOTA long-form generation huggingface/transformers#27658

I thought it would be interesting to share them to be implemented in whisper.cpp if they are not already supported.

The equivalent on the original Whisper implementation by OpenAI was implemented in a fork here.