Feat: Support Models with Reasoning Traces

[trebedea]

Description

Add support for reasoning models that output reasoning traces in the response (e.g. DeepSeek-R1).
Current implementation strips out the reasoning traces in the output parser for each executed Guardrails task, but different output parsers can be added to use the information in the reasoning trace as well (e.g. implement a custom rail that analyzes the reasoning traces).

Sample model config for DeepSeek-R1 is:

models:
  - type: main
    engine: deepseek
    model: deepseek-reasoner
    reasoning_config:
      remove_thinking_traces: True
      start_token: "<think>"
      end_token: "</think>"

Using other platforms for serving the model, such as NVIDIA AI Endpoints, is supported as long as the model name contains deepseek in the name.

MR includes:

• Documentation updates
• DeepSeek-R1 prompts and sample config
• Some refactoring in llmrails and self-check rails related to updates for using reasoning models
• Support for deepseek and google-genai as LLM providers

Related Issue(s)

• Responds to Deepseek-R1 decendants and Nemo guardrail compatibility #948

Checklist

• I've read the CONTRIBUTING guidelines.
• I've updated the documentation if applicable.
• I've added tests if applicable.
• @mentions of the person or team responsible for reviewing proposed changes.

[github-actions]

Documentation preview

https://nvidia.github.io/NeMo-Guardrails/review/pr-996

[trebedea]

@mikemckiernan , if you have any feedback about additional documentation changes (including examples).
Should we mention anywhere in the docs that reasoning models induce a serious latency overhead?

[cparisien]

This looks good, though I'd appreciate a second reviewer as well. I expect we won't have big issues since you've done plenty of "adapt to this LLM" before.

[Pouyanpi]

Suggested change

	async def test_1(rails_config):
	async def test_reasoning_traces_removed_from_responses(rails_config):

[Pouyanpi]

LGTM! Thank you Traian!