2. modify model_selection_precomputed so that all results are written into memory and then to a file at last section of code, in case that on cpu/disk seperated systems the IO takes too much time.
3. correct utils.floyd_warshall_numpy function. DONNOT use the last version.

2. correct an error in the common walk kernel. DON NOT use the old one.
3. improve the method to construct fully-labeled direct product graphs, much faster for sparse graphs.