Made kernel debug info more robust

Tries more hard to ensure that all generated functions have a
DISubprogram since otherwise they will be considered as
(nodebug)-functions by the LLVM codegen.  Also in functions with debug
data,
all calls to inlineable functions had to have !dbg.

SPIR-V LLVM translator wraps the original kernel function to make it
callable from other kernel functions for the entry point (which is
not callable). Since it didn't generate DISubprogram for the new
entry point function, it was considered nodebug. Now debug data
from SPIR-V input works better.