Add some more developed training examples. (#199)

* Use contiguous tensors for variables.

* Sketch the mnist example.

* Start adding the reduce ops.

* Renaming.

* Refactor the reduce operations.

* Bugfix for the broadcasting vectorization.

candle-core/src/op.rs

@@ -29,6 +29,8 @@ pub(crate) enum Op {
         add: f64,
     },
     Sum(Tensor, Vec<usize>),
+    Max(Tensor, Vec<usize>),
+    Min(Tensor, Vec<usize>),
     ToDType(Tensor),
     Broadcast(Tensor),
     Exp(Tensor),
@@ -354,3 +356,10 @@ impl UnaryOp for Relu {
         v
     }
 }
+
+#[derive(Debug, Clone, Copy, PartialEq, Eq)]
+pub enum ReduceOp {
+    Sum,
+    Min,
+    Max,
+}