[exp] Lazyload for multimodal inputs

[hiyouga]

What does this PR do?

Fixes #5308
Fixes #5331

In this PR, we moved the image processor's process step from the pre-processing phase to the training phase, thereby saving disk space for cached datasets. However, it could affect training throughput if the CPU is poor.

Moreover, we introduced a preprocessing_batch_size argument to control the batch size during the pre-processing phase, to prevent process hangs.

To use dataset streaming for multimodal datasets, the best practice is:

dataset: your_dataset
buffer_size: 128
preprocessing_batch_size: 128
streaming: true
accelerator_config:
  dispatch_batches: false

dispatch_batches: false is necessary for multimodal datasets since the batch size of images may differ from that of input ids.

https://huggingface.co/docs/transformers/main_classes/trainer#transformers.TrainingArguments.accelerator_config

Before submitting

• Did you read the contributor guideline?
• Did you write any new necessary tests?

[aliencaocao]

Is there anyway to restore this behaviour? Like an option somewhere?