Gradient might be doubled causing doubled learning rate

[bobcao3]

From https://github.com/taichi-dev/taichi/issues/8101

We might want to fix this to make it hyper-parameter wise compatible with torch solutions