CUDA: Don't use typemap for function name and line number

[gmarkall]

Following discussion around #7704 and #7709 it becomes clear that we shouldn't really rely on the type annotation for the function name and line number in CUDA's compile_ptx function. We should remove the use of the type annotation for this and instead get them from the Python function.

We will also need to add a test that the type annotation is present when compiling in the CUDA target, to ensure it is available for display (e.g. with numba --annotate-html).