* Use absolute error instead of relative error.
* Test on well-conditioned matrices.
* Do not repeat the same test g_repeat times (bug fix).
* Correct diagnostic output in matrix_exponential.cpp .
The latter is more stable. This fixes one of the issues with the test.
Also, make typedef's in MatrixFunctionReturnValue public; this is
necessary to get the test to compile.
The idea is that these test MatrixFunction::swapEntriesInSchur(),
which is not covered by existing tests. This did not work out as
expected, but nevertheless it is a good test so I left it in.
This is in order to get better code coverage.
Test matrix_function_3 now fails regularly because ComplexSchur
reaches the max number of iterations; further study needed.
