In the fine-tuning stage of llava, if we apply deepspeed zero3, it will put placeholders in model parameters instead of initializing real ones.

As a result, the naive load_state_dict raises errors when the code tries to load the pretrained mm projector from somewhere like mm_projector.bin.

This PR solves this by detecting if deepspeed zero3 is applied by the is_deepspeed_zero3_enabled() from transformers, and wraps the loading code with deepspeed.zero.GatheredParameters to make it effective in that case.