CI: disable Boost.Test alternate signal stack on Alpine - #11065

So according to the boost.test issue this shouldn't be enough because it comes too late. I both retried the Alpine action here and in the other PR a few times and on both came up green repeatedly. Maybe Alpine changed something on their end, maybe GH did and now it doesn't fail anymore, but in any case I'd like to hold off on merging this until we can have a clear AB test.

Maybe Alpine changed something on their end, maybe GH did and now it doesn't fail anymore

IIRC the tests were sometimes failing and fixed themselves at some point after a few retries. So given that this PR mentions AVX-512/AMX, it might be the case that the tests failed depending on the machine type the runner VM was running on (as in some have/expose AVX-512, others don't). So if that's indeed the cause, this could help to diagnose this locally (otherwise, one could add something like cat /proc/cpuinfo to the job, retry a few times over a few days and see if there are differences in there and if they correlate with CPU features).

So if that's indeed the cause, this could help to diagnose this locally

The LLM (living itself in an Alpine container) already did that for me by faking the environment and also came to the conclusion that the fix in this PR isn't sufficient in that case because the alt stack is already set once before the environment option (or command-line argument) is even parsed.

One thing that does work is compiling the alt stack out entirely with a preprocessor define BOOST_TEST_DISABLE_ALT_STACK, but that only works when using the include version of the library, which we currently only do when the shared/static versions aren't available, so to allow that override we need to add a bit of CMake code to provide a way to force the usage of the header library and pass the define through to our test targets.

@jschmidt-icinga I'm not exactly sure how to read your comment and whether it implies that you were able to reproduce that error locally.

@jschmidt-icinga I'm not exactly sure how to read your comment and whether it implies that you were able to reproduce that error locally.

1. I have a small shim library that returns ENOMEM on any call to sigaltstack() that tries to replace the existing stack, otherwise forwards the call.
2. Then I load it with LD_PRELOAD and observe testbase failing with the same message we observe in our GHA (doesn't really matter if running actual tests or running test discovery).
3. Then I use the environment variable from this PR to try to disable the alt stack for boost test but it still fails immediately with the same error we observe on the GHA.

Under the assumption that specifically AVX-512 capable GitHub runners fail, and always fail deterministically, not just sometimes, this reproduces the issue exactly and the PR does not fix it as reproduced via this method.