Inferring context-dependent computations through linear approximations of prefrontal cortex dynamics