Linear Regression Possible Improvements

Thanks again for making this!!

Just had a few things about the linear regression example.

I noticed that you're computing the gradient by manually going through the data points. After looking back at my README, I realized this was done by another contributor. I think he was trying to save memory for the intermediate nodes. Not sure what your thoughts are about this, but I think if a user can load an nxp matrix of data, we can afford to allocate O(n) extra memory for the autodiff. In that case, we can do something simpler (untested code):
```
auto expr = ad::bind(ad::norm(y-ad::dot(X, theta)));
```
(I really regret naming it norm when it's really "norm-squared"...). This should only allocate 2*(2n + 1) extra doubles in total for the intermediate nodes.
I think we can also pretty easily generalize the dimension of theta. We can just take p = theta_hat.size() for example.

Maybe a simpler implementation to replace: https://github.com/eddelbuettel/rcppfastad/blob/dbcbcd17073f620de7cb646b5df42a853d6f2de6/src/linear_regression.cpp#L23-L49 with

    const auto p = theta_hat.size();
    Eigen::VectorXd theta_adj(p);
    theta_adj.setZero(); // Set adjoints to zeros.

    // Initialize variable.
    ad::VarView<double, ad::vec> theta(theta_hat.data(), theta_adj.data(), p);

    // Create expr. Use a row buffer to store data. Then we only need to manipulate data when
    // looping.
    auto expr = ad::bind(ad::norm(y - ad::dot(X, theta)));

    // Differentiate and get loss
    auto loss = ad::autodiff(expr);

eddelbuettel / rcppfastad

Linear Regression Possible Improvements #1