[Refactor] Acclerate training based on MMEngine :rocket:

[HAOCHENYE]

@biubug6 Thanks for your excellent work on the PyTorch implementation of RetinaFace, and I believe this project has helped many other developers study and apply RetinaFace 😄 .

With the continuous development in the deep learning field, more and more training techniques and mature training pipelines have greatly improved training efficiency. OpenMMLab's new generation training framework, MMEngine, provides many out-of-box tools and high-performance training pipelines, and its standardized training process has greatly enhanced the scalability of projects.

Recently, I simply refactor this project based on MMEngine and found significant improvements in training efficiency. If you are willing, I would like to create a pull request of the refactored code to a new branch 😃. This will help community developers customize their functions more easily based on a more efficiency baseline.