Wrong per-atom stresses when running OpenMP parallel

[pastewka]

Stresses are just summed naively to the same memory location (from separate threads). There needs to be a per-thread buffer (like for forces and per-atom potential energies).