scripts: twister: Fix Twister Unit tests on Windows

[LukaszMrugala]

Unit tests were failing on Windows, indicating that current Twister code is not truly multiplatform.

Problems detected:

• Problem: Paths were deemed the main culprit. Windows has a limit of 260 character file paths by default.
  • Fix: This commit introduces a new TPath, that aims to fix the Windows limitation of supporting only up to 260 char paths by default via a path prefix disabling this limitation.
• Problem: Windows and Linux have a limit of 260 character filepath element.
  • Fix: TestSuite name formation was updated, so it wouldn't create extremely long path elements as its name, given that TestSuite name is used in dir creation.
• Problem: os.kill with a Linux-specific signal is used.
  • Fix: Process is used instead.
• Problem: pathlib's WindowPaths were not serialised correctly in the reports.
  • Fix: PosixPath serialising fix was extended to other path classes.

Unit tests were adjusted (and some disabled, as they didn't make sense on a given OS).

[stephanosio]

Problem: Paths were deemed the main culprit. Windows has a limit of 260 character file paths by default.
Problem: Windows and Linux have a limit of 260 character filepath element.

So you are basically trying to solve the same problem that --short-build-path (see #41930) already solves, but in a different way.

Does this actually work? Because Ninja and Zephyr SDK Windows toolchain binaries are not exactly long-path friendly, relying on the "ANSI version" of the I/O functions.

[LukaszMrugala]

I've had to make some fixes before answering you (Zephyr test failures). Currently, this change gives us a PASS on all Zephyr tests and all Unit Tests relevant for Windows (So without i.a. jobserver and QEMUHandler). I've tested that on a private CI.

[KamilxPaszkiet]

I tested these changes on Win 10. Everything seems fine to me.

[PerMac]

It's hard for me to evaluate this PR. I don't understand the need for all this changes to paths and in particular the new TPath class. IMO TPath adds another layer which can complicate simple operations that were done with paths. I am aware about the first 2 bullet points (char limit on Windows) but I was sure we managed to resolve it with introduction of --short-build-path. In Nordic we have a CI which is building tests/samples on Windows with twister and it works. I think I also had to add --no-detailed-test-id at some point to remove paths from test ids but cannot recall what was the root cause for this.

[LukaszMrugala]

I have gotten back to this PR now and am able to remove the DNM label.

While Twister itself can currently be run on Windows with the use of specific Twister flag, how can someone making a change to Twister know that their change is not breaking for Windows? To verify that, we ought to have a working CI which built and runs Twister tests for Windows and does a check on PRs which change Twister in any way.

We do not have such a workflow here, but I've conducted a few test workflow run in a private repo. All unit tests pass on both Windows and Linux after those changes. Without them, unit tests fail on Windows. With Unit Tests failing on Windows, we cannot have a Windows-specific CI runs and thus cannot in any way safeguard future Windows correctness.

This is not a singular PR, but a step on the way towards better Windows support.