Increase robustness of clustering partitioning

[stephenswat]

Writing this down here as a suggestion for any students or anyone else wanting to get started on traccc.

The clustering algorithm relies on being able to partition the hits into segments which are separated by at least one full row (or column) on a 2D pixel-like detector of zero-activation cells. This guarantees that there are no cross-partition clusters. The algorithm uses shared memory which is of limited side; the maximum partition size $n_\text{max}$ determines the amount of shared memory used and, as a result, the performance of the algorithm: as $n_\text{max}$ increases, performance decreases. However, this gives us an algorithm with a probabilistic success rate. For a hit density $d$, a module width or height $n$, the success probability for a given partition is approximated by $p = 1 - (1 - (1 - d)^n)^{\lfloor\frac{n_\text{max}}{dn}\rfloor+1}$. Although this chance is tiny, it still exists.

There are to projects here. First, the success probability can be increased by making the partition algorithm smarter. Second, there needs to be some mechanism to rescue the clustering in the unlikely event that a partition fails to be created.

Increasing the success probability can be done using the knowledge that a full empty row is actually a bit excessive; in reality, we only need to ensure that there is no cluster sharing between two adjacent rows. We can verify this by reifying adjacent rows and checking if they overlap under an 8-adjacency rule. This will lower performance, but probably not by much. Additional kudos if you can come up with a robust estimate of the success probability under this new rule.

Secondly, we need some logic to allocate memory in order to finish the clustering if we have an oversized cluster. This can be done fairly easily by allocating some scratch space from the device. You can allocate global memory in kernels using malloc; although this is not recommended for performance reasons, the overhead should be acceptable for this extremely rare edge case. The memory should be used to salvage the partitioning and then be deallocated.