Created by: qili93

Inherit from https://github.com/PaddlePaddle/Paddle-Lite/pull/3775

Differences:

1. SubgraphEngine still use "block_desc", not "program_desc"
2. Limit code changes in NPU only, not change to "core/program" or "core/mir".

Keeps:

1. 优化了subgraph:: Engine的执行逻辑，现在更加简单清晰。
2. 完善了npu的离线保存和读取模型的相关功能，可以作为其它需要此功能的硬件的参考
3. 新增utils：文件读写，计算md5；优化utils：split string

Remains to do: (test_shared_memory_tensor will fail after fix) 修复hiai-tensor buffer和lite-tensor buffer共享后会导致崩溃的问题。(hiai-tensor buffer共享给lite-tensor buffer，需要注意只可以由hiai-tensor去释放，而且会随着hiai-tensor的init而自动释放)