This also removes the need to compute the persmat and saves some memory
from the ViewInfos struct. This is needed to allow multiview support.
Initial testing found no major performance regression during vertex
heavy workload.
Test file:
Results:
| Platform | Master | Split Matrix |
| Linux + Mesa + AMD W6600 | 48 fps | 47 fps |
| Macbook Pro M1 | 50 fps | 51 fps |
| Linux + NVidia 1080Ti | 51 fps | 52 fps |
| Linux + Radeon Vega 64 | 25.6 fps | 26.7 fps |
Increased precision when far from origin:

