fix: avoid crash w/ remove maint flag; fix: reserve extra vram with high vram models

[tazlin]

• fix: avoid crash w/ remove maint flag if worker didn't exist

  • The worker is only registered after the first job pop. Prior to this change, if the worker operator happens to have this set when changing worker name, the worker would crash and exit on start up. As this isn't really an unrecoverable scenario (we can just skip exiting maint mode in this case), I've added a try/except to warn on whatever did happen but to continue anyway.

• fix: reserve extra vram with high vram models

  • This makes it so if high_memory_mode is on and there are big models (flux, cascade), a total of 6gb of VRAM will be reserved (per thread) to try and limit the possibility of CUDA OOMs.