fix: prevent pipeline destruction on every prompt when smart memory is disabled

ComfyUI's unload_all_models() passes 1e30 to free_memory(), which
triggered our VRAM hook to destroy the pipeline after every single
prompt execution. This forced expensive ~25-30s pipeline re-creation
on every run. Now we filter out these blanket "free everything" calls
while still correctly unloading when other models genuinely need VRAM.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>