CUDA out of memory error on datashift / dartsort spike detection

[oliche]

On the Churchland lab re-run, around 6% of recordings error as one process runs out of memory.
Ideally the memory allocation and number of jobs could handle this.

Make sure the process is stable even in presence of many spikes.

https://docs.google.com/spreadsheets/d/1uFJPW5D-Ldt4XoU7gk578iV1vfuArqVrsR_Nz7kR1to/edit?gid=1687005561#gid=1687005561