Optimize backward kernels of Sub and Add in simple cases

Context

Optimization of Sub and Add in case there is no broadcasting (gradient 1 and 0 have the same dimensions).