[NPUW] Whisper: WA when model is stuck (#34175)
When model starts to repeat itself, the same token is generated and
KVcache is overloaded. Not to crash in this case, force return eos_token
and continue processing next chunk. It's a WA solution
### Tickets:
- E-200274
