* fixed cpu mask rcnn+preparation for gpu
* fix-limit gpu memory to 30% of total memory per worker
Co-authored-by: NNikita Manovich <nikita.manovich@intel.com>