feat(merge):add merge py for llava ckpt

[tongruiliu]

What

Adds src/dataflex/utils/llava_v1d5_ckpt_merge.py, a self-contained CLI for producing LLaVA-1.5 checkpoints that evaluation engines can load.

Transformers 5.x exports LLaVA's CLIP vision tower with flat parameter names (vision_tower.embeddings.* / vision_tower.encoder.), while vLLM and Transformers 4.x evaluation stacks expect the classic nested layout
(vision_tower.vision_model.). Loading such a checkpoint fails with:

ValueError: There is no module or parameter named 'vision_tower.embeddings' in LlavaForConditionalGeneration

The new utility wraps llamafactory-cli export and then rewrites only the vision-tower key names; every tensor value is copied unchanged, so no numerical change is introduced and the training checkpoint stays untouched.

Validated on a real LLaVA-1.5-7B merged checkpoint: the converted output is byte-identical to the previously validated vLLM checkpoint (4/4 shards match sha256, 686 keys, 391 nested vision tensors, ~31s).