Long-range actors should be script interface objects

[jngrad]

Many electrostatic and magnetostatic methods implemented for the GPU have the following design:

IF ELECTROSTATICS and MMM1D_GPU:
    cdef class MMM1DGPU(ElectrostaticInteraction):
        cdef shared_ptr[Mmm1dgpuForce] sip
        cdef EspressoSystemInterface * interface

        def __cinit__(self):
            self.interface = &EspressoSystemInterface.Instance()
            self.sip = make_shared[Mmm1dgpuForce](dereference(self.interface))

        def _deactivate_method(self):
            dereference(self.sip).deactivate()

The ScriptInterface framework provides a convenient interface to store a shared pointer to a core class. It has many advantages:

• no need to manage a shared pointer, this is done by the script interface
• no need to store a raw pointer to the EspressoSystemInterface global variable
• no need to re-declare (and maintain) the core class in the .pxd file
• the Cython class can be converted to a regular Python class

The CPU implementations can also benefit from such a refactoring:

• no need to duplicate class declarations in the .pxd files
• all electrostatics and magnetostatics core global variables can be removed

Roadmap

Rewrite electrostatic and magnetostatic methods as C++ classes

• improve interface segregation and encapsulate algorithms in classes
• do dependency inversion (e.g. dipoles.hpp should include all magnetostatic methods, instead of all magnetostatic methods including dipoles.hpp)
• extract common code (P3M data structs and utility functions, FFT code)
• move actor source files to the electrostatics or magnetostatics modules
• remove the Actors framework in the core
• split electrostatic methods from magnetostatic methods
  • the core should have three new folders: electrostatics, magnetostatics, p3m (for common P3M code and FFT)
  • the core should no longer have folders actors, electrostatics_magnetostatics
• split ScaFaCoS electrostatic methods from ScaFaCoS magnetostatic methods

Write the script interface

• write ScriptInterface classes for electrostatic methods
• write ScriptInterface classes for magnetostatic methods
• write ScriptInterface classes for ScaFaCoS methods
• sort out issues related to MPI errors when throwing exceptions on worker nodes, e.g. in the constructors and tuning functions (see Script interface object constructors are not thread-safe #4449)
• sort out issues related to MPI deadlock when mixing boost::mpi communication with MpiCallbacks communication

Roll out changes to the core

• remove all electrostatics/magnetostatics global variables in the core
• add boost::optional<boost::variant<Method1,Method2,Method3>> magnetostatics_actor global variables in the core
• implement boost visitors for the energy, force and pressure calculation
• make DLC a stand alone actor instead of an extension to break dependencies in the core and lift ambiguities in the script interface (see ELC misclassified as electrostatic extension #2155)
  • e.g. what should be the state of DipolarP3M when removing the DLC "extension"? The DipolarP3M force and energy kernels are not const, therefore the actor will be altered and no longer be in sync with the DP3M python class. Making the DP3M actor an argument of DLC removes the ambiguity. In addition, the Cython Actors class no longer has to keep track of active actors vs. active extensions.
• clarify when on_coulomb_change(), parameter broadcasts and long-range actor sanity checks are really needed; right now they are called in too many places (e.g. sanity checks are run at each integration step during tuning) and parameter broadcasts override each other when an actor is adapted by ELC or DLC. Parameter broadcasts need to be removed entirely by making the tuning code synchronous, sanity checks should only be executed when an actor is active, and the on_coulomb_change() should only be called when activating, deactivating or modifying the state of an actor.
• runtime errors triggered by long-range actor sanity checks are queued and their handling is sometimes deferred to the integration loop, which is too late (fixed by Improve exception handling and fix various regressions #4479)
• break dependencies in the core: at the moment, coulomb.hpp is indirectly included in around 60 files, half of which are totally unrelated to electrostatics; to break this chain of includes, we need to make the BondedCoulombSR kernels non-inline (i.e. remove its #include "coulomb_inline.hpp" and move it to a .cpp file) and split the pair criteria into individual classes
• decouple magnetostatics events from electrostatics events (for example, modifying the DipolarP3M parameters should not reinitialize the P3MGPU parameters)