Fixes #2557.

• Extend tools/scripts/print_failures.py to understand the XML produced by --xunit-xml-output. This detects .xml versus any other extension and acts accordingly. Like the JSON codepath, success is silence. Note that a "testcase" element can have 0 or 1 "failure" elements; looping with for failure_elem was the most elegant way to handle this optionality. I checked lit's generation logic to make sure that I wasn't missing any potential forms of output (AFAICT, both unexpected failures and unexpected passes are reported with "failure" here.) The CRLF replacement needed for JSON doesn't appear to be needed for XML.
• Remove dead logic - publishArtifact was disabled a year ago due to Azure Pipelines: 11 TiB of Pipeline Artifacts is excessive #1653.
• Add a new publish step, that will upload test-results.xml. This is set to happen only if a previous step failed() (but not if the run is cancelled).
  • I've verified that this uploads artifacts for both compiletime and runtime failures, and they're easily downloadable as ZIP files. I've also verified that no artifacts are uploaded when a shard succeeds.

We'll still be able to rely on the error output that's immediately visible in Azure Pipelines for most failures, but if we encounter lengthy output (in scenarios like #2449, or for header units where the command lines are extremely long), then the XML output will be available for easy investigation.