Valgrind was complaining about using unitialized values in some places...

notbuilt) and building

Right now, some build modes reuse the same build stage value across
several actual stages. This patch splits the notbuilt stage into
notAllocated and building, avoiding that ambiguity.

- Introduce new exception BCRSMatrixError and use it throughout
  BCRSMatrix
- Introduce new exception ImplicitModeOverflowExhausted, which is thrown
  by compress() for the implicit build mode. Also improve the error
  message for that exception to really help the user understand the
  problem signalled by this exception and how to remedy the situation.

This patch adds checks to make sure the user has actually set valid
parameters before trying to allocate the matrix in implicit mode.
We use a negative value for the overflow fraction as an indicator for
this.

As part of the cleanup, this patch also makes sure that avg and
overflowsize are always initialized to a well-defined value.

This was obviously never tested - don't try to assign the overflow
fraction to the map holding the overflow entries...

We iterate over all rows anyway, so avoid the additional loop.

compress() uses iterators that are off to enable a *(++it) = ... idiom,
which already confuses me right now. That confusion will only be worse
when I look at the code again in half a year, so let's clean that up
right now.

ImplicitMatrixBuilder now has some more documentation and a second
constructor that takes all required arguments for setting up a matrix in
implicit build mode. So you can now create an ImplicitMatrixBuilder
directly on top of a default-constructed BCRSMatrix.

- Don't needlessly specialize the wrapper, we can still do that
  we we encounter the need for it.
- Add some doxygen documentation.
- Clean up formatting.

This one slipped through the cracks when renaming the build mode...

The sparsity pattern of the matrix is defined implicitly while
assembling the matrix.

call the matrix "Matrix" instead.

The given implementation only makes sense for a BCRSMAtrix, thus it should
only be used for such. This is in line with ISTLs general coding style.
Also added bit of doc.

This wrapper allows code to use the BCRSMode in implicit
build mode with the default operator[][] API during the
initial build stage.

Mpi flags were not added to pamgtest if parmetis was not present.
This is of course unwanted behaviour. The patch adds pamgtest to
the mpi tests (in addition to being part of the parmetis tests)

The config.h entry for ENABLE_UMFPACK should be in the same module
as the test for UMFPack, to avoid future conflicts with multiple
modules running the test for umfpack.

It is not needed here and does not compile with g++-4.4

The current ISTL test system relies heavily on test problems generated
with setupAnisotropic2d() from dune/istl/paamg/anisotropic.hh . The code
there was written in a fashion that forbid the use of an allocator different
than std::allocator on the BcrsMatrix.

This patch fixes this issue by giving the matrix type as template parameter
to the matrix setup instead of blocksize and fieldtype. This also makes
code easier to read. All (hopefully) tests in ISTL are adpated to the change.

Restores the old twolevelmethodtest and adds a new one to test multithreaded
computation. Fixes the autotools build system to run the multithreaded tests
too. Furthermore introduces multithreaded test with superlu/umfpack.

There was a typename used outside of template which caused
failed compilation of twolevelmethodest (at least for gcc-4.4).
This patch removes the bogus typename.

This was a left over from  resolving the conflicts from rebasing the
feature/twolevel branch.

This patch fixes a memory leak when using AMG as a coarse level solver
for the two level method. As there was no need to store the solver as
a pointer (it was copied anyway), we now simply store it by value.

[twolevelmethod][bugfix] Copy coarse level solver for thread safety. (2nd tparam of TwolevelMethod changed!)

Previously all copies of the twolevel method used the same
coarse level solver. In the case of AMG this resulted in wrong
results as it stores internal data across several applications.
This patch now always copies the internal coarse level data.

In addition the left hand side given to AMG as a coarse level solver
is copied to use it for the post method in the constructor.

Note that the meaning of the template parameters of TwolevelMethod is
changed. Its second parameter is now the type of the coarse level
solver rather than the coarse level linear operator. The latter is
deduced from the previous now. This was needed to supportt copying of
the coarse level solver.