[NVCC] Modify template for FP8 type casting to be compilable in C++20

[jushg]

Context

As mentioned by #1770 , we are having some issue with compiling NCCL v2.27.5 using C++20 (it's compilable on C++11 --> 17). We figured out the root-cause is because of this section of device code in reduce_kernel.h, which does not work in C++20. My understanding is that in C++20, the rules for aggregate initialization and implicit conversions have become stricter.

In other words, this line return toPack(VecB(fromPack<VecA>(a))); relies on the ability to construct VecB from a VecA (or vice versa) via an implicit or aggregate conversion. In C++20, this is not allowed unless there is an explicit constructor or conversion operator.

Fix Detail

I attempted to create a localized fix for this particular reduce_kernel.h code snippet here by doing explicit and manual element-wise conversion between the types.

Update:

Seems like I was being a bit too manual, just need some pushing for compiler to use the correct operator, alternatively can just do return toPack((VecB)(fromPack<VecA>(a))); \ as well --> would similarly invoke the explicit operator https://docs.nvidia.com/cuda/cuda-math-api/cuda_math_api/struct____nv__fp8x2__e5m2.html#_CPPv4NK15__nv_fp8x2_e5m2cv6float2Ev

Have compiled successfully on C++14, C++17 and C++20.

Would appreciate some help with verifying whether the fix make sense for this particular case, or if not, what would be the correct fix in order to compile with C++20.