Add next test: internet GPU pooling with speculation
Add 70B benchmark results (1.86x speedup), document same-family requirement
Fix Qwen3-72B -> Qwen2.5-72B-Instruct (Qwen3 doesn't come in 72B)
Add Pipeline Parallelism page, update Architecture with Unraid coordinator