Replace VNL with Eigen in CorrespondenceFunction::ComputeUpdates

In a large dataset, the initial iterations are dominated by VNL matrix multiplication.

e.g.

```
gramMat = points_minus_mean.transpose() * points_minus_mean;
```

```
vnl_matrix_type Q = points_minus_mean * pinvMat;
```

Eigen is significantly faster and should be used instead.