CUDA: debug=True conflates checking for exceptions and generating debug info

[gmarkall]

From the docstring of the cuda.jit decorator:

debug: If True, check for exceptions thrown when executing the kernel. Since this degrades performance, this should only be used for debugging purposes. Defaults to False. (The default value can be overridden by setting environment variable NUMBA_CUDA_DEBUGINFO=1.)

This also enables the generation of debug information. Ideally there should be two separate flags, one for debuginfo, and one for checking for exceptions. Perhaps:

• debuginfo: Set to True to enable debug info generation.
• exception_check: Set to True to enable exception checking.

with debug=True implying that both of the above are True - this will provide consistent behavior with existing versions of Numba as well as providing more fine-grained control over debugging functionality.

[gmarkall]

There is some alignment / improvement needed with the rest of Numba here (e.g. debug switching on boundscheck for the CPU target) so I'm removing the good first issue label as it requires some alignment / thinking about with Numba.

[gmarkall]

This won't make it into 0.55, optimistically bumping to 0.56.

[dlee992]

I feel this feature would be useful if we have. I will try to work on this and post feedback if I met some obstacles.

Anything I need to know before hacking? gentle cc @gmarkall

[gmarkall]

I have this prototyped somewhere already, let me see if I can find it...

[gmarkall]

It's in this branch: gmarkall/numba@cuda-decorator-refactor...gmarkall:numba:cuda-compilation-refactor

Which is based on this PR: #8773