Remove runtime recursion from find_ntuplets()

[makortel]

From cms-sw/cmssw#35473 and cms-sw/cmssw#35542 for all other versions than hip (that was done in #233).

[makortel]

@fwyzard @VinInn For Alpaka I went with this instead of the specialization for DEPTH == 0 because partial specializations of functions are not allowed ("partial" caused by the additional T_Acc template argument). By quick test the throughput improvement is similar order (3-4 %) than in cuda without caching/async allocator and in kokkos (that both use the specialization as in the original PR). If you have any better suggestions, let me know.

[fwyzard]

After seeing it, I actually like the approach with if constexpr better than the one with the specialisation for 0, as it keeps things more localised.

We should check that it works well also for the native CUDA and HIP cases.

[makortel]

Would it be ok for you to do that in a subsequent PR?

[fwyzard]

Sure.

[makortel]

I opened an issue to remind about that #240.

[VinInn]

I consider "partial specializations of functions are not allowed" a defect in Alpaka.
The use of "if constexpr" instead of template specialization to terminate recursion is a more general coding pattern far more reaching that the case in hand

[fwyzard]

I consider "partial specializations of functions are not allowed" a defect in Alpaka.

Well, maybe a defect of C++ :-/ ?

[VinInn]

let's rephrase: The need to specifically template in an intrusive fashion ALL functions with T_ACC is a serious issue in Alpaka

[fwyzard]

We don't need to follow the same approach in our code, but I don't think a template parameter is a bag choice.
For example, after working on #241 I think that a template parameter is easier to deal with that a different namespace - especially if we don't want to multiplicate every symbol, object, etc.