Generate a report from a CI testing run

[ev-br]

In both numpy-fork and array-api-strict, we generate the report from a separate CI run, so that a test suite is run twice: once for testing and another time for the report.
It feels natural to run it only once, and generate the report from the testing run.

This, however, would run into several hiccups:

• the test runs are typically done with --skips-file= or --xfails-file since the testing signal is "CI passes modulo known failures"; to handle xfails the reporting would need to parse xpasses from the pytest output.
• some test suites are run with skips instead of xfails via the environment variable, since generally xfails are expensive with hypothesis; so a robust solution could be to run with skips, then rerun the skips specifically (maybe with a lower number of hypothesis examples), and merge the two reports.

All in all, this might be more trouble than it's worth.

[kgryte]

So long as running with a skips file or xfails file doesn't interfere with the reporting (i.e., a skipped test still reports as skipped and a test failure still reports as a failure), then it is up to the array library how it wants to do things in CI.

The demo.yml is a demo, not prescriptive. Given that array libraries have varying CI setups and practices, I would expect that one or more would choose to run CI differently. The main thing is that array libraries provide an honest accounting of their compliance. How that happens is up to them. They only need to ensure that the report is valid according to the report schema.