eig/eigh

[mrader1248]

If the library cusolver is chosen in the eig function of linalg, it behaves like a function I would call eigh. Wouldn't it be clearer if a function eigh is added and the cusolver-code from eig is moved to eigh? In addition a cula-code for eigh could be added in the future.

[lebedov]

Since numpy provides both eig and eigh functions (and since cusolver doesn't provide generalized eigenvalue decomposition yet), this suggestion is a good idea.

[jalfonsi]

Could you please add support also for tridiagonal matrices like eigh_tridiagonal taking as input only arrays for diagonal and bands (see scipy for example)?