Shift Parallelism: Low-Latency, High-Throughput LLM Inference for Dynamic Workloads