VectorTree operations in mrcpp

[gitpeterwind]

The present "OrbitalVector" class, could be transferred to mrcpp (as TreeVector then, possibly with other name) and upgraded. The goal is that:

• MPI parallel operations can be performed on the new class. We have already many such operation used in practice: rotation, dot, addition (exchange?). This will make a clearer separation between the methods which are mpi collective and the ones which are individual.
• The data content can be stored in several ways, and flags show which ones are used (can be stored in several ways simultaneously). At present we already have many ways the data is stored: directly, as my_orb(), shared memory, in Bank as orbitals/functions, in Bank as nodes. This will also give better control over memory usage, which is an important bottleneck.
• The storage of data content is adapted transparently: if two consecutive operations are using local/node type operations, there is no need to transform the orbitals back and forth.
• The methods/operation are not performed when asked(!), but only when the results are required. Instead a list of implicit operations is stored. When the result is really required, all are performed at once. This will allow for example that several transformations (rotation and potential multiplication for example) can be done on the same node, while it is in cache (perhaps only L3, but possibly L2?). (This part needs to be thought of more in details)
• Many (all?) operation can be performed based on a local (node) approach. A local approach is more efficient from a cache point of view and easy to parallelize. This is at present underexploited in mrchem.
• the density/dft scheme should also fit into this scheme.
• Bank moved from mrchem to mrcpp
• Complex trees (=two real trees?)