Optimize CI workflow. (#1743)

See pull request for details. Summary:

* Reduce dependencies between CI jobs, allowing more of them to run in parallel.
* Use Linux for as many jobs as possible.
* Merge matching "-ndll" and "-samples" jobs. This way, the latter doesn't have to waste time recreating the former's environment, and the workflow summary is simplified. Downside: this delays package-haxelib.
* Update to a newer macOS version, as the old version massively slowed down Homebrew.
* Skip the HelloWorld test for Mac and iOS, as those two can build slowly under some circumstances, delaying package-haxelib. These could be restored later, if performance seems good enough.
* Add `HAXE_VERSION` environment variable, so future updates will only have to make one change. Unfortunately, doing the same for the Windows/Mac/Linux versions would severely hurt performance, so those remain hard-coded.
* Split up some steps within jobs for better organization.