This fixes a crash when a key.pkl file is present in a compilation cache
directory, but no .dll (or .so)

disabled the c code of the mod operator as it don't give the same result at the py version when their is one or both of the signe with an operand.

This alias lets user force host-allocation of float32 matrices even when device
is gpu

output type, which should never happen...

Little hack in reshape() function to set the tag.shape of the output if the new shapeis a tuple or list.