[VLM] reduce CPU peak memory in multimodal tensor hashing#21123

yhyang201 merged 6 commits intosgl-project:mainsgl-project/sglang:mainfrom

yhyang201:reduce-hash-peak-memyhyang201/sglang:reduce-hash-peak-memCopy head branch name to clipboard