Fix some bugs that cause minor annoyances…

[savetheclocktower]

…during the build process.

Claude ferreted these out; the precipitating annoyance was the fact that yarn install in the Pulsar repo often somehow forces me to re-run the download-libiconv step. Claude doesn't know why the existing version gets removed during that process, but says that yarn build doesn't catch the omission because the are-we-still-fresh metadata doesn't know to check for libiconv.2.dylib. The fix is to move that file into the same directory as superstring.node.

Also, Claude spotted some bugs in the fetch-libiconv-61.sh script.

The proof here will be in the CI; if this works just as well in CI, then I'll consider it to be a lateral move at worst. These bugs would hardly ever surface during Pulsar builds because we're always starting from scratch, but fixing them may mean I have slightly fewer headaches.

[savetheclocktower]

Since this PR was open, it became the testing ground for a bug I uncovered in superstring. It's possibly the reason why the Windows editor tests are crashing on ~40% of runs! This bug was pretty severe in the sense that it was possible to trigger a crash entirely from the JavaScript layer, so I didn't want to put it in another PR and queue it up behind this one.

While testing that bug, I also noticed that the test:native task wasn't actually doing anything on Windows and was vacuously reporting a zero exit code. Once I actually got those tests to run on Windows, it was clear that they had never been designed to run on Windows — C++ types and idioms that don't work with the standard Windows C++ toolchain, for instance. Those were annoying to troubleshoot because I don't have a Windows dev machine locally, so it was more like pushing and waiting for CI to tell me what was still broken.

Now that this is green, I'm gonna land it soon and then see if it fixes the Windows crashes (which, luckily, seem way more likely to happen in CI than in real life).