TODO

known bugs:

* Using normalized test and trial spaces in the uniform grid version
  of ASTI does not work, as it would need to use different spaces
  for different directions.
* When calling the update() method of normalized spaces, we do not
  update the gridviews associated with the GridViewFunctions in the
  coefficients of the inner product that handles the normalization.
  That is not a problem for constant or piecewise constant coefficients
  but might lead to wrong evaluations for other GridViewFunctions.
* src/kernel_svds.cc does not work for non-wavelet kernels due to
  changes in the scattering kernel functions.

janitorial tasks:

* ASTI: check that I handle the RHS correctly.
  IIRC I handle scattering, bv, rhs etc all in one vector
  represented in the test space basis. This probably has to change.
  We might need to handle it in two different vectors represented in
  different bases.

potential performance improvements:

* Using the Eigen::Map wrapper around Dune's matrix and vector classes
  might spare us some temporary copies in the Cholesky decomposition
  and in the wavelet code.
  We need however to check that Map works with vectors of FieldVectors.

proposed features:

* Implement a multi-grid preconditioner so that we can replace the
  UMFPACK solver with a CG solver.
  See e.g. Roberts and Chan 2017
* In the ASTI code we can currently only use scattering kernels that
  are uniform in space. For more complex benchmark problems it would be
  useful if we could use different scattering kernels in different
  regions of the spatial domain.
* The current interface of refined spaces, introduced in commit
  bd31710ea5390cfed3c423cf150d9ee475880015, is really ugly and could
  probably improved by creating a new class SubElementLocalView
  used for objects that get generated by the localView. This way the
  localView is only bound to an element and can be a const reference
  again, while the SubElementLocalView handles all the subelement
  specific stuff.