fix: leave slot filling to only dialog service using nlu (PL-1109) [bugfix]

[trs]

By having slot filling in the NLU and Dialog services, we're effectively calling the NLU up to 3 times a turn. The Dialog service is already responsible for slot filling, so lets keep it there and remove it from the NLU service.

Slot filling in the Dialog service was also not retaining intent context between re-prompts, meaning you could get in an infinite loop of re-prompts as it would "forget" the previous entity. This has been fixed.

Prediction in the Dialog service now always uses the NLU setting, regardless of the project, since the LLM does no slot filling.

Filtered intents and entities are added to prediction during slot filling. Support for these isn't quite there at the NLU level, but the Performance team is actively working on that.