[CIR][HIP|CUDA] Generate global storing CUDA|HIP stub function

[koparasy]

On HIP when launching a kernel we pass as a first argument a global variable that points to the device stub function.

We follow OG design by having a map that pairs globals to symbols. In CUDA this is effectively a nop, as CUDA passes the device stub as a first argument.

[AdUhTkJm]

This PR suggests we should use the function names, rather than the FuncOp itself, as the key to kernelHandles. That is because the same function might be created twice, once with incomplete types and once with complete types.

[koparasy]

This PR suggests we should use the function names, rather than the FuncOp itself, as the key to kernelHandles. That is because the same function might be created twice, once with incomplete types and once with complete types.

Ok, I will add this.

[koparasy]

This PR suggests we should use the function names, rather than the FuncOp itself, as the key to kernelHandles. That is because the same function might be created twice, once with incomplete types and once with complete types.

Ok, I will add this.

Done

[bcardosolopes]

Nit but otherwise ready to go