inotify_rm_watch() generates IN_IGNORED events into the inotify
queue, which immediately triggers another watch_dispatch() call,
which calls inotify_rm_watch() again, generating another IN_IGNORED
ad infinitum. This pegs the CPU and makes input laggy on every
config reload.
Fix by removing inotify_rm_watch() entirely. inotify_add_watch()
alone handles both cases:
- Same inode: modifies the existing watch (no event generated)
- New inode (delete-and-recreate): creates a new watch; old watch
on the unlinked inode generates a single IN_IGNORED (one-shot,
not a loop)