[Docs] Warn users when NLP service does not offer constrained sampling

Motivation

A number of NLP adapters do not offer constrained sampling, sometimes referred to as Structured Outputs. When building agents with Parlant and these services, JSON Mode would continue to be used, and may return invalid or even empty responses.

Solution Proposal

1. Document the NLP adapters that offer full structured outputs functionality once the relevant FRs, and possibly #678, are completed.
2. Add constrained sampling functionality to the forthcoming Emcie NLP service
3. Review the remaining NLP adapters and add a warning that they may provide inferior quality responses to their respective documentation pages.

NLP Adapter Status

Adapter	Support?	Link
OpenAI	✅	#679
Mistral	✅	#682
Anthropic	✅	#684
Gemini	✅	#687
Azure	✅	#688
AWS Bedrock	✅	#691
LiteLLM	✅ [1]	#692
Qwen	✅ [1]	#695
OpenRouter	✅ [1]	#696
Cerebras	✅	Reference
Fireworks	✅	Reference
DeepSeek	❌	Reference
Ollama	❌ [2]	Reference
GLM	❌ [3]	Reference
Huggingface	N/A [4]	Reference
Snowflake Cortex	❌	Reference
Together	❌	Reference
Zhipu	❌	Reference

Notes:

1. Model-dependent
2. Can be resolved in #678
3. Duplicate of Zhipu adapter (#697)
4. Adapter is for embeddings only