Reducing GPU Offload Latency via Fine-Grained CPU-GPU Synchronization