Fix or work around broken tests in new analysis

[cameel]

See also #14510 (comment)

Remaining fixes before we can have green CI and merge the prototype:

• Patch codegen issues enough to make tests pass in experimental analysis #14660
• Assert json artifacts input/output when using experimental #14659
  • Remove unimplemented assert for fetching linked object #14726
• Exclude function dependency graph tests from ast import export #14722
• Exclude experimental error codes from error_codes.py #14727
• Consistent bytecode report output between CLI and StandardJSON in presence of UnimplementedFeatureError.
  • IMO we should fix CLI and Standard JSON do not generate the same outputs in case of errors at code generation stage #13925, but there's no consensus that this is the best approach and working around this in prepare_report.py will probably be quicker to do.
  • It was a bug: Fix prepare_report.py ignoring compiler exit code #14731

[ekpyron]

We definitely should not waste effort on #13925. Why doesn't the filter in prepare_report.py already cover these cases?

[cameel]

I tried hard to avoid prepare_report.py having to parse CLI output and fish ICEs out of it. Instead it just checks whether it can find the artifacts. No artifacts == error. The filter is only necessary for Standard JSON to ignore inconsistent artifacts.

But actually, turns out that this was not the cause. The script had a bug - it was not looking at the exit code, which for CLI is another indication that something went wrong. The script originally only handled Standard JSON and there the exit code does not matter. When I was adding CLI support later, I must have forgotten to add a check for that.