Implement same flavor index per warp on GPU

[Qubitol]

Currently, events processed by the GPU may have different flavor index per threads in a warp.
This can result in thread divergence, especially when some HELAS routines are zero for some flavor combinations.

This has been solved on the SIMD point of view, by grouping events having the same flavor index together in the same SIMD vector.

For GPU we would need to do the same, and ensure the warps of 32 threads have all the same flavor index, to favour perfect lockstep processing.

[theoheimel]

Probably it is best to ensure that most of the batches only contain one flavor, but still allow for mixed batches rather than introducing padding. For example: if we have 48 events with flavor 1 and 48 with flavor 2, it will probably be better to have 3 blocks (one flavor 1, one flavor 2, one mixed) rather than 4 with padding (two flavor 1, two flavor 2).

[Qubitol]

Yes, I agree, it's a good compromise, coding will be easier, and it's not really a matter of a single warp having divergent threads that will bring performances down.