Add the gradient for reduce-sum. (#162)

* Add the gradient for reduce-sum.

* And add the gradient for the broadcast ops.

* Add some backprop tests.

* Add some linear regression example.

candle-nn/src/optim.rs

@@ -39,7 +39,7 @@ impl SGD {
         let grads = loss.backward()?;
         for var in self.vars.iter() {
             if let Some(grad) = grads.get(var) {
-                var.set(&var.sub(&(grad * self.learning_rate)?)?)?
+                var.set(&var.sub(&(grad * self.learning_rate)?)?)?;
             }
         }
         Ok(())