[CIR][CUDA|HIP] Removes special handling for CUDA|HIP global function

[koparasy]

The GetAddrOfFunction returns now the host address.

[koparasy]

@AdUhTkJm I think we should remove the special handling over here. Earlier you had this but that never evaluates to true on cuda, cause lang.HIP && lang.CUDA always is false when building nvCuda. On #1333 I was rigorous and checked whether the function is a definition.

This removes the special case completely and continues to CodeGen, in which you fill in the global body with the stub.

[AdUhTkJm]

The OG says whenever we refer to the device stub in HIP, we need to use the handle address instead. So I think this special case is indeed necessary, and should be implemented along with the generation of handles.

It was my fault to write langOpts.HIP && langOpts.CUDA. It should be the change that you suggested in your previous PR, (langOpts.HIP || langOpts.CUDA) && hasAttr<GlobalAttr> && IsForDefinition. The same branch is present in OG.

[koparasy]

Just letting you know that I am working on this.