Fix deprecated call

[ClementPinard]

I noticed that the tutorial was updated but not the codebase, so here it goes. It is now on par with tutorial http://pytorch.org/tutorials/advanced/cpp_extension.html

change AT_ASSERTM by TORCH_CHECK change .type() by .scalar_type() change PackedAcessor to PackedAccessor32

This hopefully fixes #65 , #66 (although 1.6 norally only return deprecation warning)

Second significant change : change fminf and fmaxf to their fmin and fmax counterparts, make sur that the right template is used by casting the 0.0 to scalar_t . This is probably not needed anymore as it was probably a bug with nvcc and gcc7 but it might help people with old configs get the grad_check working.

This fixes #27 and #42