decide which part of interface has to be optimized ( probably nothing until we implement reader for HDF5, XML or large Raw data files, and try to use the same input interface for input of large data)
then make inlined only neccessary functions and carefully move as much as possible into accessors.cc including explicit instantiation of support classes. This should speedup compilation of the code that use the accessors.
implement operator -> without allocation (shared_ptr), i.e. put Accesors into Iterators Create corresponding accessor at construction of the iterator.

better tests:
- common set of quantities with different kind of values (scalar, vector, tensor, discrete, ..), common points and elements for evaluation
- for individual Field implementations have:
  - different input
  - possibly different EPETCT_EQ tests, but rather have majority common

make FieldPython dummy class if we do not have python, so that we need not optional code elsewhere

vyresit navaznost na lin sys - solve a export seq vektoru, redukce ... ?
inv_a - predava se pri konstrukci, ale neumoznuje jeji reuse - aktualizaci assemblace resp. nutno si na ni drzet ukazatel venku ... reseni ?
? remove old_4_new - just for LSView
automatic preallocation
eliminated block given by IS
in place Schur
? nemodifikovat puvodni system, leda skrze jeho metody

START_GLOBAL_TIMER(tag) - this calls the start_timer, which creates local timer on the correct place in the hierarchy, further this timer is added to the list of global timers, this contains groups of timers with same tag, and collect/sum data from these timers in the report.

remove transport_sources
in create_transport_matric_mpi, there there is condition edge_flow > ZERO this makes matrix sparser, but can lead to elements without outflow and other problems when there are big differences in fluxes, more over it doesn't work if overall flow is very small