perf: recalibrate vram_plan (leave GPU headroom) + refresh perf README

Update model card, app.json (patch size + VRAM plan) and inference configs