Use create_new instead of atomic-file-write

[mattfbacon]

This provides the same functionality but without temporary files, platform-specific code, fragility of O_TMPFILE support, and an extra dependency.

[abonander]

Unfortunately this implementation is more fragile than the one it replaces, as it does not consider race conditions, or even what happens on the first run before the file exists.

If two crates in the workspace have the same query, they'll fight over each other trying to delete and then re-create the file if Cargo decides to compile them concurrently.

[mattfbacon]

Good points, I forgot to explicitly handle these conditions as non-errors. I think the latest commit should handle these properly.

[jplatte]

I'm still not convinced that avoiding the dependency is worth it, but I guess there's nothing obviously wrong with it anymore. Could work 😅

[abonander]

LGTM, just needs to handle PermissionDenied as discussed above.

[mattfbacon]

OK