Refactor MD setup for pre-parallel cleanup

[Audrey-777]

Summary

This PR performs a scoped MD code refactor in preparation for future parallel optimization. It does not intend to change
functionality, numerical behavior, or performance.

Changes include:

• Extract MD runner creation into a local factory helper with std::unique_ptr ownership.
• Move DP/NEP model temporary buffers to solver members for clearer lifetime management.
• Add MD kinetic/stress state helper structs and pure helper functions while preserving existing APIs as wrappers.
• Introduce shared MD test fixtures and remove duplicated test source registration.

Validation

• Built production target successfully:
  cmake -S . -B build-prod -DBUILD_TESTING=OFF -DENABLE_MPI=OFF -DENABLE_LCAO=OFF -DUSE_ELPA=OFF
cmake --build build-prod -j2

• Ran targeted MD tests successfully:
  ctest --test-dir build-mpi-test -R 'MODULE_MD' --output-on-failure

All targeted MODULE_MD tests passed.

[mohanchen]

Great work. Feel free to proceed with your planned optimizations. One reminder: MD simulations consume significant memory for large systems, so try to minimize array usage. In the final testing phase, please use systems with a larger number of atoms.