it appears some refactoring is necessary on run_share4v.py

thank you for your work. i am using it for captioning images. i didn't get a chance to review it all, but noticed a few issues. i'm exemplifying one of them below.

https://github.com/InternLM/InternLM-XComposer/blob/f31220eddca2cf6246ee2ddf8e375a40457ff626/projects/ShareGPT4V/share4v/eval/run_share4v.py#L26

- this function accepts args however parameters such as `max_new_tokens`, `temperature` are hardcoded.
- this file does not return the output but prints it.
- do_sample is true even if temperature=0.
- this is a prompt adherence issue: the model does not follow instructions such as `limit number of words to 50`.
- num_beams is useless as it results in `AssertionError: Only support batch size 1 (yet)`