Mark one test from mostly-non-JS/JVM backends as platform-specific behaviour, test on macOS

[huonw]

This goes through many backends and marks one test as platform_specific_behavior to ensure the backends have some testing on platforms other than x86-64 Linux (macOS, and arm64 Linux).

The particular value in doing this is validating the set-up code, e.g. got the right configuration for installing/downloading the tool, rather than validating the details of the behaviour. I think we generally assume most tools behave similar across platforms, once they're up and running.

This PR focuses on non-JVM and non-JS backends, meaning raw binaries (ExternalToolm TemplatedExternalTool) and Python tools installed from PyPI (PythonToolBase, PythonToolRequirementsBase). It particularly focuses on the backends that don't require installing any additional software in CI, see #21620 and #21621 for additional backends that do require installing Go and Docker respectively.

The impression I get is the JVM and JS tools are more platform-independent, often? For instance, their lockfiles/artifacts don't mention specific platforms, unlike Python wheels.

I found these backends via the following shell command, plus other exploration:

for file in $(rg --files-with-matches "class.*(PythonTool.*Base|ExternalTool)" --glob '*.py' . | sed 's@/[^/]*$@@' | sort | uniq); do
  rg -q "platform_specific_behavior" $file || echo $file
done

I may've missed some!

This increases the time to run tests in CI on the relevant platforms, by slight more than 2×.

platform	before	after
Linux x86-64	runs all tests always	no change
Linux arm64	~3:30	~7:30
macOS x86-64	~5:30	12:00 - 13:30
macOS arm64	not run in CI (#20993)	no change

Related: #21608, #20193, #20993

[benjyw]

Part of me thinks we should run all tests on all platforms and be done with it...

[huonw]

Yeah, agreed: #20193 😄