This can interfer drastically with the ongoing parse jobs. Most
notably, I just encountered a case where a long-running parse job
was ongoing, but all other parse threads where waiting since the
cleanup thread already tried to lock for write, thereby starving
the parse jobs.

I think we should wait with the cleanup until the parse jobs have
finished. To do so, we try to lock for write and give up if this
does not succeed.

Seems to work as I intended. On shutdown, we still wait for the jobs
and cleanup unconditionally. But while a big project is parsed, no
cleanup is taking place.