Add multi-image input to GRPO trainer

[Andcircle]

The current solution seems only support single image input, also
the image processing logic here https://github.com/huggingface/trl/blob/main/trl/trainer/grpo_trainer.py#L1363
is not that flexible. Maybe should enable customized data_collator

In order to support multi-image input, I think VLLMCLient also need to be updated:
https://github.com/huggingface/trl/blob/main/trl/trainer/grpo_trainer.py#L1441
Is it possible just use vllm.LLM, each prompt is a conversation (including images)