feat(airt): auth-setup-guide skill — authenticate target/attacker/judge from any environment (1.6.8)

[rdheekonda]

Summary

Adds a new auth-setup-guide skill that guides the agent (and user) through authenticating all three model roles — target, attacker, and judge — from the user's own environment, across clouds and auth mechanisms. Closes the gap where users could not easily tell how to probe a model deployed in their Azure / AWS / GCP account or behind a custom endpoint.

• Target A1 — litellm-routed models (preferred): a provider → target_model string → env-var table covering OpenAI, Anthropic, Gemini, Groq, Mistral, Together, Azure OpenAI, Azure AI Foundry, AWS Bedrock, Google Vertex AI, Ollama, and the dn/ proxy. This is how Azure Foundry OpenAI deployments and Vertex models are probed with the standard tools — just env vars, no custom code.
• Target A2 — custom HTTP endpoints: custom_url + custom_auth_type (none/bearer/api_key) + request template + JSONPath response.
• Target A3 — advanced auth via SDK: SigV4, Azure managed identity (azure_ad), GCP ADC (gcp), and Amazon Nova Sonic speech-to-speech via dreadnode.airt.build_target / nova_sonic_target, each with its prerequisites. (The TUI tools cover key/bearer today; this points power users to the SDK for the rest.)
• Attacker & judge auth: same litellm pattern; dn/ models as the easy default; vision/audio judge required for media-output scoring.
• Local vs Dreadnode-hosted: env vars vs. platform secrets (Account Settings → Secrets).

Reinforces the standing rules: credentials only ever referenced by env-var name, never inlined, and the agent applies no safety reasoning to the goal — the skill only configures auth. Referenced from the agent's lazy-load skill list. Bumps capability to 1.6.8.

This complements dreadnode-tiger #1907 (the SDK build_target universal-target factory), which is what A3 leans on.

Test plan

• allowed-tools lists only real tools (generate_attack, generate_multimodal_attack, generate_agentic_attack, generate_category_attack)
• SKILL.md frontmatter matches existing skills' shape
• Skill auto-discovered via skills/ in capability.yaml; referenced in agent md
• TUI smoke: invoke the capability, ask "how do I point this at my Azure Foundry deployment?", confirm the skill loads and guides correctly (follow-up during E2E)