Fuse inner loop kernels in device CKF

[stephenswat]

The inner loop of the device CKF consists of five loops: material interaction application, measurement counting, candidate finding, hole writing, and propagation. I believe that the middle three can be easily merged into a single kernel, reducing the amount of work we have to do on the host and simplifying the code a lot. This commit makes that change.

[krasznaa]

While not understanding most of the changes in the logic, I like the general direction quite a bit. As long as the code still works after this in the same way as it did before, I'm very much on board. 😉

[beomki-yeo]

Reducing the number of kernels is definitely good. I think the CPU finding algorithm will also need to remove orig_param_id and skip_counter to keep the symmetry but this can be done later.

Have you checked the performance of this?

[stephenswat]

PR updated.

Have you checked the performance of this?

It's essentially identical:

Before:

Reconstructed track parameters: 3061231
Time totals:
                  File reading  80234 ms
            Warm-up processing  984 ms
              Event processing  8227 ms
Throughput:
            Warm-up processing  98.4611 ms/event, 10.1563 events/s
              Event processing  82.2799 ms/event, 12.1536 events/s

After:

Reconstructed track parameters: 3063909
Time totals:
                  File reading  79725 ms
            Warm-up processing  27888 ms
              Event processing  8229 ms
Throughput:
            Warm-up processing  2788.83 ms/event, 0.358573 events/s
              Event processing  82.2956 ms/event, 12.1513 events/s