lora adapter hot swap inference

Size GPU memory for multi-LoRA serving in vLLM and SGLang: adapter slots, max_lora_rank padding, KV cache trade-offs, swap latency and tenancy rules.