Optimisations to scan kernel

[RAMitchell]

No description provided.

[RAMitchell]

Some performance numbers:

dataset	max_depth	PR #152	main	This PR
Airline	8	1.240454	1.227512	1.232933
	16	1.699671	1.534235	1.481347
Wine	8	1.045785	1.042129	1.042439
	16	1.855366	1.678566	1.528984
YearPredictionMSD	8	3.023942	3.040562	2.881305
	16	27.307748	26.455262	22.803253
adult	8	2.034721	2.713914	2.613187
	16	3.818123	2.341759	2.211143
covtype	8	6.403403	6.805757	6.589184
	16	19.547341	18.109158	16.711339
epsilon	8	24.606926	21.542766	21.221759
	16	430.767659	320.014831	255.979749
higgs	8	1.503556	1.508633	1.452655
	16	6.907475	6.856589	5.593349
kddcup99	8	1.722053	1.610332	1.695249
	16	2.735567	2.057228	2.037838
make_classification	8	18.130453	18.325938	17.865216
	16	131.233685	131.645496	114.272764
make_regression	8	4.30786	4.039117	4.103192
	16	34.079777	32.005323	28.106314
mnist	8	18.576233	13.380752	13.331378
	16	193.952795	128.222687	110.471264
satimage	8	1.724287	1.533877	1.50608
	16	6.63816	5.858919	4.212776