Use selfadjointView when computing eigenvectors

[lisitsyn]

• Fixes wrong PCA results with dense eigensolver (thanks William Curram
  for reporting!)

[lisitsyn]

@iglesias want to take a look? ;)

[iglesias]

I don't have context on this one but I say, sure and go ahead.

[lisitsyn]

@iglesias there was an issue: when computing PCA with arpack we have all good. But when it comes to 'dense' eigenmethod which is just eigen3 thing, eigenvectors are like (0 0 0 0 1 0). I took a look and apparently we use selfadjoint view when computing covariance, thus resulting in a upper triangular matrix. With the fix it treats this matrix as a self adjoint, ignoring the lower triangular part (but diagonal).

[iglesias]

Ok, thanks for the explanation!