[VLM] optimize VLM processing

[nanjiangwill]

followup #1232

1. Remove extra processing of multimodal train inputs
2. Add CI for VLM FSDP
3. Update --rotary-base for VL model

[zhuzilin]

I think that it's a.bit hard to understand the meaning of multimodal tensor, could you rename the multimodal tensors to something like multimodal train inputs and still calling the orginal field as multimodal inputs.