fix: Cancel ongoing build scripts/workspace discovery on exit

[ChayimFriedman2]

Flycheck and proc macro server process handles are stored in the GlobalState, using wrappers (JodChild or CommandHandle) that kill the process on drop. So when we exit and GlobalState is dropped, they're automatically cancelled.

However, build scripts and workspace discovery are triggered inside thread pool tasks. When GlobalState is dropped, they're left untouched, and when r-a process exits and the handles are destroyed, the processes are detached.

The way I went to solve this is by adding an option to cancel all tasks on a thread pool, by injecting a panic into it. Of course you cannot force a thread to panic (this has the same problems as killing a thread), so instead the thread must check some state periodically. Most of our tasks are have short execution time before executing the next database query, so we can cancel the database, making Salsa inject a panic. Except those that invoke child processes.

So I went and changed JodChild to also check periodically for cancellation. This is non-trivial because the standard library has neither "read with timeout" nor "wait with timeout" primitives for pipes, so we're forced into platform-specific code. I also introduced a JodCommand wrapper around Command to make it harder to use waiting operations by mistake, and added Command and Child to clippy::disallowed_types.

This was still not enough, because the extension itself has not delivered the shutdown event correctly, and also had a too short timeout for it (200ms). So I changed it to send a shutdown request in deactivate(), and increased the timeout to 2s.

I have no idea how to test this automatically, but I tested this manually by setting a build command that waits forever then closing VSCode. Before those changes, the build process is still there; after them, it is killed properly.

Fixes #22732.