feat: backend watchdog

[mudler]

Is your feature request related to a problem? Please describe.
Issues like #1333 makes the backend completely unusable as it stalls entirely and does not allow any other request to go in. Besides, this is particularly problematic for me as I use grammar to generate data :)

It currently requires manual/automatic intervention to restart local-ai if it gets into that condition. And as this is outside of our control, we need to take a solution that works generically.

Describe the solution you'd like
A watchdog inside LocalAI that detects backend stalls (configurable, enable/disable it) and shuts it down automatically

Describe alternatives you've considered
Didn't considered other alternatives yet

Additional context
N/A