faster featurizer

[djsutherland]

The old Cython featurizer only took two minutes on low1 once dummies had been created; this new one takes two hours. Dunno how long dummies took, but not two hours.

What's the significant difference? Speed that up....

[djsutherland]

One thing that might help is to call MKL straight instead of through numpy: gemm would save a dot-then-scale, sincos is much faster than np.sin plus np.cos (numpy/numpy#2626 (comment)), probably some other savings to be had in rff_embedding (or use fastfood).