ndarray: porting Python's autograd (#274)

Ref: https://github.com/apache/incubator-mxnet/blob/065adb3702c110af7b537799be3ec9c16c27a72b/python/mxnet/autograd.py

* API ported
    * attach_grad
    * grad
    * mark_variables
    * get_symbol
    * record
    * pause
    * train_mode
    * predict_mode
    * backward

* An example

```julia
x = NDArray([1 2; 3 4])
mx.attach_grad!(x)

y = mx.record() do
  mx.square(x)
end

mx.backward!(y)

mx.getgrad(x)
 # 2×2 Array{Int64,2}:
 # 2  4
 # 6  8
```

NEWS.md

@@ -75,6 +75,8 @@
 
 ### `NDArray`
 
+* A port of Python's `autograd` for `NDArray` (#274)
+
 * `size(x, dims...)` is supported now. (#TBD)
 
   ```julia

src/MXNet.jl

@@ -105,6 +105,7 @@ include("broadcast.jl")
 
 include("ndarray.jl")
 include("random.jl")
+include("autograd.jl")
 
 include("name.jl")
 include("symbolic-node.jl")