Optimize Tensor-scalar operations by not relying on brodcasting

[marty1885]

The current way Etaler handles Tensor-scalar operations is by supply constructors that converts scalar values into Tensors of shape {1}. Then rely on broadcasting to process the entire Tensor. This works and is only a few lines of code. But is very slow (Have to read from memory every operations, can't put the scalar in a register). We'll need another solution to make this fast.