scikit-learn: Adding fused types to Cython files

The current implementation of many algorithms, such as Stochastic Gradient Descent, Coordinate Descent, etc. only allow input with float64 and int64 dtypes due to the adoption of Cython fused types may result in explosion of the generated C code. However, since scikit-learn has removed Cython files from the repo and re-generate them from every build, it provides a good chance to refactor the ".pyx" files by introducing Cython fused types. This will allow those algorithms to support float32 and int32 dtypes data, which is currently casted into float64 and int64 respectively, and therefore reduce the waste of memory space.