Draft: Memory optimisation of the parameterised Kalman filter

This MR is an incremental update to the parameterised Kalman filter that should have no* impact on the output of the Kalman filter and should only create a speedup of the algorithm by decreasing the amount of registers used and thus allowing for better utilisation of the GPU.

The changes were presented in WP2.

RE: Pipeline throughput tests All the throughput test results look fishy to me. All but one if the sequences that run in the full set are without this algorith but show significant decreases on 2/3 of the tested GPUs. The small throughput test show 5%-10% improvement, which is also larger than anything I ever managed to reproduce on the testbench.

• I snuck in a small adjustment to the derivates of the correction to the UT->T extrapolation, where high order corrections were dropped but they were not removed from the derivatives.