Long-Context Attention from Kernel Efficiency to Distributed Context Parallelism