Row and column extractions on 2D tensors should use fixed views

We currently use dynamic views while extracting rows and columns of a 2D tensor

~~~c++
Tensor<double, 3, 4> A;
A(0, all)  --> TensorViewExpr<A>
~~~

this make it unnecessarily impossible to construct tensors of the right shape at compile time when we already know this info which is `Tensor<double,4>` in this case. 