added float4 and double2

[WillTrojak]

Added jit'ing for float4 and double2 vector types in the CUDA generator.

[FreddieWitherden]

Do you have any benchmarks for these? It will increase compile times so I want to be sure that we do see a benefit somewhere.

[FreddieWitherden]

Not sure if we need to bother with the alignment.

[WillTrojak]

Here is the FP64 performance improvement in % for N=48^3

p	mat	hex	pri	tet
2	m0	-3.533371763	4.843308571	1.9023938
2	m3	-0.295154775	0.297843008	5.808760296
2	m6	-1.418440302	-0.049529643	-0.071124285
2	m132	-0.097657075	-4.690991397	11.03387621
2	m460	-1.240993204	0.059032363	2.869200715
3	m0	8.185044182	1.21527681	0
3	m3	0	-2.477560435	-0.112825122
3	m6	-1.608576253	0.772438562	-0.080645839
3	m132	-1.10576634	1.523184537	3.296704216
3	m460	-0.029065774	0	0.058428688
4	m0	-1.746725526	-0.626782048	1.411388864
4	m3	0.653198534	0	1.183432388
4	m6	-0.139722193	5.780346028	0.009443318
4	m132	-0.512157025	0.557107275	-0.012451606
4	m460	-1.4669897	-0.008854826	0
5	m0	0.008564597	-0.011397145	-0.766288992
5	m3	1.480361034	0.222750352	-0.922382174
5	m6	-0.720272118	-0.297150933	-0.326509939
5	m132	-0.807297097	0.494840479	-1.078167653
5	m460	0.510315992	12.28417455	-0.014238971
6	m0	0.1712243	0	0.040380291
6	m3	10.2580566	-0.13422927	-0.007912969
6	m6	0.118645903	0.049778555	0.006923897
6	m132	0.036737528	5.175943762	0.004347303
6	m460	0.494356025	-0.675444632	-0.00839014

[WillTrojak]

It's not clear to me why the performance would go down; it could be noise.