Full fine-tune vs LoRA vs QLoRA on one L40S: Qwen2.5-1.5B, memory, speed and held-out loss, measured