Text recognition runs three jobs at once at night

At night and over the weekend the server's transcription service now works on up to three recordings at the same time instead of one, so it finishes about 4 hours of recording per hour instead of 1.6. By day it stays at one job. The museum app keeps priority: the service only gets CPU that nobody else needs, holds back further night jobs while the CPU is under pressure, and stops a job of its own before memory runs short. A job stopped that way is queued again automatically.