[SGLang CI] Deselect a GPU-only VLM test (#50913)
[SGLang CI] Deselect test_colliding_low_hash_bits_do_not_reuse_image_kv (needs a GPU)
It fails on our CPU runner with "Cannot access accelerator device when none is available"
(empty_gpu_cache()). SGLang registers test_vlm_input_format.py for its CUDA CI only.
Co-authored-by: ydshieh <ydshieh@users.noreply.github.com>