Learning Sampling Distributions Using Local 3D Workspace Decompositions for Motion Planning in High Dimensions