Seeing the Plumbing: Profiling vLLM Speculative Decoding on NVIDIA Blackwell