Use template instead of duplicating code in elementwise_util.h (#9058)

Now all the apply functions share a common implementation, which means
further changes (e.g., parallel_for, generating specialized dtypes for
the case where all inputs have the same type) don't need to be repeated
3 times.

(Interestingly, this seems to increase the effectiveness of the
following parallelization change. Not entirely sure why, but I checked
the generated code for optimized op_where and it seems to have improved,
which is surprising.)