init_kv_cache probed comfy_kitchen.flash_attention_decode_is_available()
unconditionally, which calls torch.cuda.get_device_capability() and raises
ValueError when the execution device is CPU (e.g. MiniMax Music3 text
encoder offloaded to CPU on low-VRAM GPUs). Guard the probe with
comfy.model_management.is_device_cuda(), matching the pattern already
used elsewhere in this file (ar.py's cuda_device check).
Fixes#15607
chelsealong