Add retry with exponential backoff for network tools

[longzhi]

Problem

Network tools lack proper retry logic:

Tool	Current Behavior
web_fetch_tool.rs	No retry on failure
web_search_tool.rs	No retry on failure
webhook.rs:69	Has retry but uses linear delay (RETRY_DELAY * attempt)

Transient network failures (timeouts, 5xx errors) cause immediate tool failure, which then consumes a reasoning loop iteration to retry at a higher level.

Proposed Solution

1. Implement configurable retry with exponential backoff for web_fetch_tool and web_search_tool
2. Convert webhook.rs from linear to exponential backoff
3. Only retry on timeout and 5xx errors (not 4xx client errors)
4. Make max retries and base delay configurable

Impact

• Resilience — transient failures handled transparently
• Efficiency — avoids wasting LLM reasoning loops on retryable network errors