Unify the guardrails system with the existing plugin hook infrastructure:

- Add before_request and after_response hooks to plugin types
- Extend before_tool_call/after_tool_call with richer context and
  modify capabilities (after_tool_call now returns results)
- Wire up all four hook stages in the agent runner and tool adapter
- Move Gray Swan implementation to extensions/grayswan/ plugin
- Remove guardrail registry and apply functions

The Gray Swan guardrail now follows the standard plugin pattern,
registering handlers via api.on() for each stage. Configuration
remains unchanged (guardrails.grayswan in openclaw.json).

This enables third-party guardrail plugins using the same hook API.

Add a shared base class/factory that simplifies guardrail extension
implementations by handling common hook registration boilerplate.

Extensions now implement a simple interface:
- evaluate(ctx, config, api) -> GuardrailEvaluation
- formatViolationMessage(evaluation, location) -> string
- onRegister(api, config) [optional]

The factory handles:
- All 4 hook registrations (before_request, before_tool_call,
  after_tool_call, after_response)
- Stage config resolution (enabled, mode, blockMode, includeHistory)
- Error handling with failOpen support
- Monitor mode (log-only)
- Content extraction per stage

Refactored extensions:
- llamaguard: 634 → 373 lines (-41%)
- grayswan: 590 → 394 lines (-33%)
- gpt-oss-safeguard: 550 → 311 lines (-43%)

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>