fix(core): implement network boundary security for agent endpoints

[RinZ27]

Proposed Changes

This PR implements a first-line defense against SSRF by introducing an IP denylist in the AlmanacResolver. It ensures that agent endpoints resolved via the Almanac API do not point to private, loopback, link-local, or cloud metadata IP addresses (RFC 1918).

Key changes:

• Added is_safe_url helper in resolver.py using Python's standard ipaddress library.
• Integrated the safety check into lookup_endpoint_for_agent to filter out unsafe endpoints.
• Added warning logs for filtered endpoints to assist debugging in development environments.

Linked Issues

Fixes #854

Types of changes

• Bug fix (non-breaking change that fixes an issue).

Checklist

• I have read the CONTRIBUTING guide
• Checks and tests pass locally

Further comments

While there are ongoing discussions regarding a more comprehensive security middleware and trust gate architectures (e.g., #865), this IP denylisting provides immediate protection against common SSRF vectors. It blocks access to internal resources like 169.254.169.254 and local admin interfaces by default.

Verified the logic with test cases covering:

• RFC 1918 private ranges (10.x, 172.16.x, 192.168.x)
• Loopback (127.0.0.1, localhost)
• Link-local/Cloud Metadata (169.254.169.254)
• Public URLs (agentverse.ai, google.com)