[CIR][CUDA] Generate attribute for kernel name of device stubs

[AdUhTkJm]

Now a __global__ function on host will be generated to a device stub, with an attribute recording the corresponding kernel name (mangled name on device of the same function). The dynamic registration phase will be implemented in LLVM lowering.

For example, CIR generated for __global__ void global_fn(); looks like

#fn_attr1 = #cir<extra({cuda_kernel_name = #cir.cuda_kernel_name<_Z9global_fnv>})>
cir.func private @_Z24__device_stub__global_fnv() extra(#fn_attr1)

[bcardosolopes]

Awesome, some comments inline