Fix LlamaFirewall scans inside the MCP event loop

[3nesdeniz]

Summary

• make guardrail provider checks asynchronous end to end
• use LlamaFirewall's native scan_async() API for server configurations and tool responses
• update the built-in providers, mocks, and call sites for the async contract
• add regression coverage for both ALLOW and BLOCK decisions

Fixes #39.

Root cause

mcp-context-protector executes guardrail checks inside its active asyncio event loop. LlamaFirewall.scan() starts another event loop with asyncio.run(), so both the configuration-review and tool-response paths could fail with:

RuntimeError: asyncio.run() cannot be called from a running event loop

LlamaFirewall provides scan_async() for this environment. The provider contract and both call chains now await that API directly.

Verification

• Reproduced the failure on the current main branch with the official MCP Python weather server.
• Verified configuration scanning and a real get_alerts tool call after the fix.
• Full test suite: 302 passed
• Coverage: 69% (CI threshold: 55%)
• Focused regression suite on Python 3.11, 3.12, and 3.13: 13 passed on each version
• make lint: Ruff formatting/checks and 100% interrogate coverage passed
• uv build: source distribution and wheel built successfully
• git diff --check: passed

[CLAassistant]

All committers have signed the CLA.