These are the only bit of shared state between the Request user thread and RequestManager that aren't covered by RequestHandle::stateMx. The problem was that they were not covered by anything, which meant that they were not guaranteed to be coherent between threads.
Also fix WriteFile being unable to overwrite existing files. The rename would fail because the file was still open, and the sanity remove in response to that would also fail for the same reason.
We can't rely on atexit, handlers registered with it are in a hard to establish ordering relationship with destructors of static and thread-local objects.
Namely: no, yes, and yes and ask at startup.
The install_check option is thus replaced by the can_install option. -Dinstall_check=true maps to -Dcan_install=yes_check, while -Dinstall_check=false maps to -Dcan_install=yes. -Dcan_install=no is new and is recommended for downstream packaging, where -Dinstall_check=false was historically used.
Also improve error messages about bad configuration here and there and scatter configuration code in subdirectories, where they can be closer to their areas of effect.
... while retaining all the functionality of stamps.def.
Also fix stamp names encoding only 32 bits of the timestamp, migrate from stamps.def to stamps.json if the latter doesn't exist, delete both on migration to the shared data directory, rescan stamps at startup, and make rescanning a painless process in general by removing invalid entries and adding missing entires at the beginning of the list.
Because for some reason the hdiutil convert step would sometimes fail with "image corrupt" when given an image the hdiutil create step had just previously created. Too cursed for my pay grade; I took the easy way out.
Request ownership is no longer flaky. Requests are now owned by the code that makes requests, and Requests and the RequestManager co-own RequestHandles. RequestManager disowns a RequestHandle if it's done with it or if Request code reports that it's no longer needed.
All libcurl code has been moved to RequestManager. This is nice because once NOHTTP is removed, we can add any number of RequestManager implementations, for example one for Android.
Client outliving RequestManager is still a problem, this will have to be addressed later.
More precisely, refactor the code responsible for routing these GameController events to the Lua side. The issue with the previous solution was it relied on preprocessor macros to switch between Lua-ful and Lua-less builds.
