[no squash] Fix building unittests & benchmarks by upgrading Catch2

[appgurueu]

Fixes #14692, alternative to #14693, by simply upgrading Catch2 to v3. Catch uses a separate implementation and header now, which helps resolve the issue. This uses the amalgamation.

Bonus: Catch2 version upgrade.

Small change (I consider this an improvement, but you may differ): Gets rid of #define CATCH_CONFIG_CONSOLE_WIDTH 160. I think this (arbitrarily?) chosen number is too large. It is also inconsistent with unit tests. Using the Catch default (80) here is better (and simpler) IMO.

Fun fact: You can run benchmarks via --run-unittests --test-module <benchmark> now. Catch2 also prints "Randomness seeded to: <seed>" now.

To review, look at the second commit (8b02699). The first commit updates the vendored copy, which causes a lot of noise.

How to test

1. Compile with unittests and benchmarks enabled.
2. Confirm that --run-unittests and --run-benchmarks work as expected.
3. Confirm that both still work as expected if --test-module is used.

[sfan5]

ps. Catch did not go from 20kloc to 200kloc. The git diff is largely from Catch2 test files (text files). Removing the tests would reduce the size of Catch2 by ~5 MB.

Importing 200kLOC is a terrible idea.
We should either vendor the amalgated source, require catch2 to already be installed or consider a submodule. (in order of personal preference)

Downloading content at build time is a nuisance, so FetchContent is not worthy of consideration.

• Gets rid of our "modification" to Catch which forced everything to be formatted as microseconds - now it will format as ns, µs, ms, etc. as appropriate. I prefer this.

I'm pretty sure this was required to make the automated benchmarks work correctly.

[appgurueu]

Alright, I'll go with the amalgamated source, and re-apply the patch to output everything in microseconds.

CI reveals that the return value of Catch::Session().run(...) is just an exit code now, meaning it fails when there are no tests to run. However we can simply tell catch to consider zero tests a success via the config.

[appgurueu]

Alright, switched to the amalgamation. The custom cout / cerr / clog pose a problem, because they cause ODR definition violations vs. Catch internals. We need to tell Catch to always use our redefinitions (for benchmarks as well) which just use rawstream, which is what I did now.

[JosiahWI]

Thank you for fixing the ability to build unittests and benchmarks together. This is a nice clean solution. I will be closing #14693 in favor of this.

[JosiahWI]

LGTM

[sfan5]

I also suggest changing the author of the first commit to something artificial to not skew the LoC change statistics.

[appgurueu]

I also suggest changing the author of the first commit to something artificial to not skew the LoC change statistics.

Done, thanks for the reminder. (I remember having done this initially, but it probably got lost when I switched to the amalgamation, redoing the commit.)

[sfan5]

CI green = good