Add support for np.float32 matrices in lars_path.

Some single-precision routines have had to be ported from cblas. Later
on we could probably remove most of them since new scipy has
triangular_solve function and cholesky_delete could be implemented in
Python.