feat: improve qwen2-vl startup (#2802)
* feat: tokenize each request individually and increase warmup image size
* feat: adjust rotary embed and avoid cuda graphs of size 2 and smaller
* fix: address image resize and rebase changes
* feat: update to run qwen2-vl tests
* fix: tweak param types
