TOF improvements

[KrisThielemans]

• add test for symmetric kernel of TOF (see TOF bin fix for RayTracingMatrix #1645)
• remove zeroes from the cached projection matrix. Before, for every TOF bin, we still had row with the same size as the row of the non-TOF matrix. This meant a lot more memory was used than needed, and a potentially lot of multiplications with zero.

[KrisThielemans]

@z-k-li maybe you could have a look at this

[KrisThielemans]

@z-k-li any news? Note that I suspect that this will make the first (full) iteration slower but later ones faster. It'd be interesting to check the timings reported.

[z-k-li]

@z-k-li any news? Note that I suspect that this will make the first (full) iteration slower but later ones faster. It'd be interesting to check the timings reported.

Hello Kris, Sorry for the late reply, I only just had time to test this.

The speed improvement seems big.

I found that the new PR takes about 150 s per subiteration (wall clock time), whereas the old version takes about 280 s per sub-iteration. I did not observe substantial difference between different full iterations.

Overall, for a 4i5s reconstruction of a 220 × 220 × 645 MRD85 dataset using 16 CPU cores, the new PR completes in ~70 minutes, compared with ~124 minutes for the old version.

[KrisThielemans]

That's great news! Maybe you could report memory usage here as well, but I'll merge.