Self-hosted inference orchestrators compared: LocalAI, exo, GPUStack, vLLM