Enforce LLM timeout configuration that was previously dead code

[Iron-Ham]

Summary

The Configuration struct in FoundationModelsClient defined maxAttempts, requestTimeout, and retryDelay but these values were never used. The session.respond() call had no timeout wrapper, causing the tool to hang indefinitely if Apple Intelligence became unresponsive - this is the likely root cause of #14 where users had to reboot their Mac to recover.

Changes:

• Wrap LLM generation calls in withTimeout() using Swift's withThrowingTaskGroup to race the operation against a timer
• Implement exponential backoff on retries (30s → 60s → 120s by default)
• Change default timeout from 20s to 30s for better first-attempt success rates
• Make FoundationModelsClient conform to Sendable for Swift 6 strict concurrency
• Fix unrelated test that was checking for non-existent prompt text

Test plan

• Build passes with swift build
• All 49 existing tests pass
• Manual test: run scg generate with a large diff and verify timeout behavior
• Manual test: verify exponential backoff by observing retry timing in verbose mode

Fixes #16