Marking a subprogram, variable, or location in the DIBuilder API
requires passing in a `numba.core.ir.Loc` object. This makes sense when
all debuginfo generation comes from lowering, but creates weird
dependencies when generating code that doesn't relate to a specific
piece of IR - for example, the kernel wrapper generated in the CUDA
target. To mark locations in the kernel wrapper, one must import the
`Loc` class from `numba.core.ir`, and construct instances of it to pass
in, just to mark locations in generated code (or do something more odd
with a "pretend" `Loc` object).

Since the `Loc` instances are only used for the line number, this commit
changes the interface of the relevant `DIBuilder` methods to accept a
line number instead of a `Loc`, breaking the spurious coupling.

Doing this provides a couple of improvements:

- Breaking on any kernel launch, for example with
  `set cuda break_on_launch application` now results in GDB being able
  to find the source line at the break location - this enables the
  backtrace to show the mangled function name, source file, and line
  number. It also enables the text user interface to display the source
  when the kernel launches. From this point, stepping through
  line-by-line is feasible with the `next` command.
- Breakpoints can be set on a specific function and argument types, by
  setting a breakpoint on the mangled name.

(neither of these worked prior to this change)

Issues that remain include:

- GDB doesn't demangle the names of functions, like it seems to in the
  CPU target examples. Oddly, the act of marking the subprogram seems to
  break demangling - prior to this commit, the location of the wrapper
  kernel isn't known, but it does demangle correctly.
- Most locals appear to be optimized out even with `opt=0`. This also
  needs further investigation.

Includes:

- Calls `llvm_to_ptx` once for each IR module when compiling for debug.
- Don't adjust linkage of functions in linked modules when debugging,
  because we need device functions to be externally visible.
- Fixed setting of NVVM options when calling `compile_cuda` from kernel
  compilation and device function template compilation.
- Removes `debug_pubnames` patch.

Outcomes:

- The "Error: Debugging support cannot be enabled when number of debug
  compile units is more than 1" message is no longer produced with NVVM
  3.4.
- NVVM 7.0: Everything still seems to "work" as much as it did before.
  Stepping may be more stable, but this needs a bit more verification.

Fixes numba#5311.

…vements

…vements

- Use `self.py_func.__code__` to get the filename and line number for
  the function for `prepare_cuda_kernel`, which seems a bit cleaner /
  neater that digging into the type annotation for it.
- Fix typo and explicitly link to an issue comment.