Blocked linear system

issue 198

A first step towards the block linear algebra is to switch to Thyra linear algebra structures, without introducing blocked structures yet. This will change/add some interfaces, mostly in AbstractDiscretization (added) and Application/Workset (changed), plus some internal changes in the evaluators (only gather/scatter ones, in theory). Delaying the introduction of block structures to a second step will allow for a quick(er) merge into master, so that new features can be developed with the new interfaces.

Here's a tentative todo list for the first step. If someone completes them, mark them as done.

• (a) Add getters in AbstractDiscretization for Thyra vector spaces. These can simply create the VS on the fly, wrapping the corresponding Tpetra_Map.
• (b) Replace Tpetra/Epetra structures in Workset with Thyra ones. As a half-step, we can simply add Thyra structures,
• (c) Change Application/ModelEvaluatorT, to work directly on Thyra objects, rather than Tpetra/Epetra. In particular, ModelEvaluatorT can simply forward InArgs contents to the Application.
• (d) Change gather/scatter evaluators, so that they can work with a Workset storing only Thyra objects.

Steps (b)-(d) can be performed incrementally on the objects stored in Workset. For instance, we can start from the residual evaluation only. All other structures will be left as Tpetra/Epetra.

Note: ModelEvaluator (epetra version) does not inherit from Thyra's ME, which complicates things. There is a Thyra adapter for EpetraExt ME, which would allow to use Thyra for both Epetra and Tpetra builds, but it's not used in Albany. I see two options: (1) make ModelEvaluator inherit from the Thyra adapter to EpetraExt's ME; (2) remove Epetra capabilities altogether. The first option is the most backward compatible. But the fact that Epetra-only builds are not possible makes me wonder why shouldn't we go for (2)...