Expand Developer Guide section on arch symbol visibility with complete example and past learnings

[jrhemstad]

This issue can be closed by expanding the CUB Developer Guide section on symbol visibility with simple, but complete sketch of a template free function that ultimately invokes a kernel, e.g.,

template <typename ...>
auto bar(...){
   kernel<<<...>>>(...);
}


template <typename ....>
auto foo(...){
   return bar(...);
}

but with the appropriate annotations, arch-specific guarding, etc.

It should satisfy the following scenarios:

libA.so and libB.so both include <foo_kernel> and invoke with same template arguments:

libA.so compiled with -arch=sm_80

#include <foo>
auto A(){ return foo(...);}

libB.so compiled with -arch=sm_90

#include <foo>
auto B(){ return foo(...);}

// main.cu
#include <A.h>
#include <B.h>

int main(){
   if( arch == 80 )
     A();
   else if(arch == 90)
     B();
}

Calling A() should always invoke kernel for sm_80, and likewise B() calls kernel for sm_90.

This code example should be compiled and run as part of our regular CI. It should incorporate any testing that @gevtushenko developed in https://gitlab-master.nvidia.com/gevtushenko/cub_visibility that detect problematic scenarios.

This should pull in any additional learnings from @gevtushenko original investigation in https://gitlab-master.nvidia.com/gevtushenko/cub_visibility such that this repo can be closed and the developer guide is the source of truth. This should reflect the updates in CUDA 13 with __global__ implying static.