add: support for eval set for fine tuning scripts

[jbuchananr]

Issue:

Currently, there is no way to pass an eval set to be used during fine tuning to prevent over fitting.

Fix:

Modify internvl_chat_finetune.py to take optional data set via meta.json to be used for evaluation.

Example Usage:

torchrun \ 
  ...
  --meta_path "/home/ubuntu/meta_train.json" \
  ...
  --eval_meta_path "/home/ubuntu/valid_meta.json" \
  --do_eval True \
  --eval_strategy "steps" \ 
  --eval_steps 100 \
  ...