Make cuda kernels operate on all tracks at the same time

[jorenretel]

Until now the decompression was happening in parallel and the interval_to_values operation was done on each track seperately. The main reason for that was the cupy searchsorted function operating on one "values" array at the time. Here, we have a modified version of that searchsorted kernel that does multiple arrays at the same time. After this was done it was also possible to rewrite intervals_to_values so that it operates on all tracks at the same time.

[jorenretel]

Note: I am importing some cuda kernel code here from a private variable in cupy:https://github.com/pfizer-opensource/bigwig-loader/blob/53cbe8697c773ceec91fbeeb74eaf58625a9652f/bigwig_loader/searchsorted.py#L5C1-L6C1

I am partially doing this so I don't have to copy the code... I also don't really expect this code to change any soon. Running GPU tests in GitHub Actions would immediately detect if something goes awry. But we don't have a GPU runner yet.