Add auxiliary losses directly imposed on params

[tuffr5]

The general framework for training is:
x --> hash encoding --> network --> y <<loss

If I impose auxiliary losses directly on params like

x --> hash encoding --> network --> y <<loss
           ^               ^
           ^               ^
   auxiliary loss1   auxiliary loss2

I wonder that is it possible in current tiny-cuda-nn? Currently, I tried in the way I indicated above to regularize the params, resulting in a terrible model.

[tuffr5]

It seems like that the gradient calculation on network is correct, but the hash encoding is not correct. One can do backward multiple times for different losses to get better results, but it slows down the training. Is there any solution for this?