tensor.zeros_like for elemwise ops

now it is two tests, one with an integer, asserting grad is 0
one with a float, asserting grad is what it was before

(didn't I do this before? did the rebase lose them somehow?)

gradient and an input have the same type, so the warning makes no sense

(this shouldn't change the output in any case, just eliminates
unnecessary calls to Op.grad)

throughout the grad machinery

checking

needed

with respect to float