Classification of motor imagery and execution signals with population-level feature sets: implications for probe design in fNIRS based BCI