feat: `obol openclaw set-model` for per-instance provider/model switching

Problem

Changing an OpenClaw instance's model or provider routing requires re-onboarding (obol openclaw onboard --interactive) or manually editing the values-obol.yaml overlay. There's no quick CLI verb to switch an existing instance's model or toggle between LiteLLM-routed and direct provider access.

Proposed UX

# Switch model (still through LiteLLM)
obol openclaw set-model obol-agent --model gpt-4.1

# Switch to direct OpenAI (bypass LiteLLM)
obol openclaw set-model obol-agent --provider openai --direct --api-key sk-proj-...

# Switch back to LiteLLM-routed
obol openclaw set-model obol-agent --model claude-sonnet-4-6

Context

The infrastructure already supports both routing modes:

• LiteLLM-routed: buildLiteLLMRoutedOverlay() — all traffic through litellm:4000/v1
• Direct provider: buildDirectProviderOverlay() — native API access, key in per-instance Secret

Currently only accessible via obol openclaw onboard --interactive which recreates the full deployment. A set-model verb would patch the overlay and re-sync without full re-onboarding.

Scope

• obol openclaw set-model <id> --model <name> — change primary model, keep current routing
• obol openclaw set-model <id> --provider <name> --direct — switch to direct provider
• obol openclaw set-model <id> --provider litellm — switch back to LiteLLM routing
• Re-sync instance after overlay change (helmfile sync)
• Validate model exists in LiteLLM config (for routed mode) or provider API (for direct mode)

Start by tracing the existing obol openclaw onboard --interactive flow and the buildLiteLLMRoutedOverlay() and buildDirectProviderOverlay() helpers. Define how set-model updates the per-instance overlay, handles routing and credentials, validates the requested model, and performs helmfile sync; done means all listed CLI modes work without full re-onboarding.