[LITE][NPU][XPU] Refine subgraph pass, and support NPU/XPU model generation at execution time (#2576)
