vllm: Defer disposable GLM MTP head

vLLM now supports deferred execution for the GLM Multi-Token Prediction (MTP) head to optimize inference throughput. This architectural change prevents the overhead of processing disposable prediction heads during the decoding phase, allowing for more efficient resource allocation when running GLM-based architectures.