Towards cumulative race time regression in sports: I3D ConvNet transfer learning in ultra-distance running events