scikit-learn

Small performance improvements (~10%) and it no longer uses custom dot
matrix function dot_over. Except for solve_triangular, all other
methods can operate on float32 arrays.

Thinkint out loud:

    The complexity of LARS is the same as an OLS, and looking at the
    benchmarks scikits/learn/glm/benchmarks/bench_glm.py it shows that
    we are faster than scipy's lstsq, so I don't think there is much
    more room for improvement. I'll benchmark against the R package
    and see how far we are.

In arrayfuncs: removed dot_over and added min_pos.