Support of custom allocator for atoms_positions, atoms_total_forces and atoms_new_colvar_forces

[HanatoK]

I would like to avoid extra data copying when implementing the interface for GPU-resident NAMD. Currently #783 holds an intermediate buffer for transposition, and copy the data from the buffer to atoms_positions, atoms_total_forces and atoms_new_colvar_forces. The CUDA kernel cannot write to atoms_positions, atoms_total_forces and atoms_new_colvar_forces directly because they are not allocated with page-locked memory managed by the CUDA runtime. To solve the issue, a custom allocator can be used for them:

#if defined(COLVARS_CUDA)
#include <cuda_runtime.h>
#endif

#if defined(COLVARS_CUDA)
  template <typename T>
  class CudaHostAllocator {
  public:
    using value_type = T;

    CudaHostAllocator() = default;

    template<typename U>
    constexpr CudaHostAllocator(const CudaHostAllocator<U>&) noexcept {}

    friend bool operator==(const CudaHostAllocator&, const CudaHostAllocator&) { return true; }
    friend bool operator!=(const CudaHostAllocator&, const CudaHostAllocator&) { return false; }

    T* allocate(size_t n) {
      T* ptr;
      if (cudaHostAlloc(&ptr, n * sizeof(T), cudaHostAllocMapped) != cudaSuccess) {
        std::cerr << "BAD ALLOC!" << std::endl;
        throw std::bad_alloc();
      }
      std::cerr << "CudaHostAllocator: allocate at " << ptr << std::endl;
      return ptr;
    }
    void deallocate(T* ptr, size_t n) noexcept {
      cudaFreeHost(&ptr);
    }
    template<typename U, typename... Args>
    void construct(U* p, Args&&... args) {
        new(p) U(std::forward<Args>(args)...);
    }

    template<typename U>
    void destroy(U* p) noexcept {
        p->~U();
    }
  };
#endif

#if defined(COLVARS_CUDA)
  template <typename T>
  using allocator_type = CudaHostAllocator<T>;
#else
  template <typename T>
  using allocator_type = std::allocator<T>;
#endif

/// \brief Current three-dimensional positions of the atoms
std::vector<cvm::rvector, allocator_type<cvm::rvector>> atoms_positions;
/// \brief Most recent total forces on each atom
std::vector<cvm::rvector, allocator_type<cvm::rvector>> atoms_total_forces;
/// \brief Forces applied from colvars, to be communicated to the MD integrator
std::vector<cvm::rvector, allocator_type<cvm::rvector>> atoms_new_colvar_forces;

But to use this allocator there will be pervasive changes to Colvars code where only std::vector<T> is used.

[HanatoK]

Actually after thinking again I am not quite if this is necessary. If Colvars finishes #655 this might not be necessary.

[HanatoK]

After testing, I found that it is still necessary to support allocating vectors on CUDA managed host-pinned memory for better performance. This has been done in b830dca and d8dd54f.

[HanatoK]

Done in #783.