feature: add output rail checking mode to check_async()

[m-misiura]

Did you check the docs?

• I have read all the NeMo-Guardrails docs

Is your feature request related to a problem? Please describe.

check_async() currently runs input rails only (it extracts the last user message and evaluates it). There is no way to use /v1/checks to evaluate a bot/assistant response against output rails. Callers who need to check LLM-generated content for safety before returning it to users must use generate_async() or build their own rail invocation.

Describe the solution you'd like

Add a rail_types parameter to check_async() (and expose it in the /v1/checks request schema) that lets callers specify which rail types to run, e.g.:

• ["input"]: current default, check user messages against input rails
• ["output"]: check assistant messages against output rails
• ["input", "output"]: run both

Describe alternatives you've considered

Callers use generate_async() with output rails; this works but defeats the purpose of a lightweight check-only API.

Additional context

Parent issue: #2014
First PR: #2013

[m-misiura]

cc @tgasser-nv, @Pouyanpi; if you think this is a good proposal, please assign me to it :)