This is a set of 3 changes:

1. Simplify the structure of the solver class:
   • make firstMoves and winMoves private members and add getter for them
   • replace mutex protected bool with atomic bool
   • do not directly expose m_shouldEnd, but provide a stopExecution function
   • use in-class initialization of member variables
   • remove debug variable, and the DEBUG_HINTS define which controlled it, use #ifndef NDEBUG to print the debug messages for now
   • Solver: use a unique_ptr for the MemoryManager

2. Introduce a smaller SolverInterface, hiding the complexity of Solver

This change:

• extract all the publically used Solver functionality into SolverInterface, which consists only of virtual functions
• move solverinterface into its own header, to minimize includes
• converts users which used Solver directly to SolverInterface

3. Convert Solver into a template, templated over the number of piles
   • The number of piles is known in each Solver subclass at compile time, thus a subclass with n piles does now inherit from Solver<n>.
   • m_number_piles and its setter become redundant with this change, and are removed
   • This enable the conversion of member variables which were dynamically allocated arrays into std::arrays, increasing type safety and causing an abort/exception on out-of-boundary access. Furthermore, the compiler can now better reason about some of the loops (though this doesn't really matter performance wise)

Everything still compiles, and the Solver can still solve each game