More realistic training setup. (#210)

* More realistic training setup.

* Compute the model accuracy.

* Very inefficient backprop for index select.

* More backprop.

* Fix some backprop issues.

* Backprop fix.

* Another broadcasting backprop fix.

* Better backprop for reducing ops.

* Training again.

* Add some gradient tests.

* Get the training to work.

candle-core/src/op.rs

@@ -48,6 +48,7 @@ pub(crate) enum Op {
     Binary(Tensor, Tensor, BinaryOp),
     Unary(Tensor, UnaryOp),
     Cmp(Tensor, CmpOp),
+    // The third argument is the reduced shape with `keepdim=true`.
     Reduce(Tensor, ReduceOp, Vec<usize>),
     Matmul(Tensor, Tensor),
     Embedding(Tensor, Tensor),