Get rid of global state in fail and warn

[erlend-aasland]

Here's an experiment I thought of earlier today, as a step towards error handling without global state. One of the biggest pains wrt. getting rid of the clinic global is that fail depends on it too much. So here's the experiment:

• tear of the global state from warn_or_fail
• identify the failing tests, and amend them by supplying the line number and filename explicitly

It is a crude change, but it kind of works, and I think it might be an improvement (if polished a little bit).

Side-effects:

• some errors now generate more accurate line numbers
• most warnings (the few we have) generate less accurate line numbers (or no at all)
• cpp.py now have a cleaner error interface towards clinic (IMO)

IMO, we should just use a variety of custom exceptions, catch these during parsing, format the error message and reraise it (for either the CLI or the test suite to catch).

[erlend-aasland]

After the latest yak-shaving, the diff for this PR is now greatly reduced.

[erlend-aasland]

Victor had some remarks about this PR; I'll try to address them tomorrow:

• it is unclear what happens if you pass lineno but not filename

[erlend-aasland]

• it is unclear what happens if you pass lineno but not filename

Come to think of it, that is also a concern with the current code; this PR does not alter the status quo.

[vstinner]

I wrote a different approach: python#114752

Maybe some of these changes are still relevent with my change, I didn't dig into the code yet.

[erlend-aasland]

Upstreamed as python#115510