fix(api): retry on 429 rate-limit with exponential backoff (QUA-645)

[Desperado]

Summary

Free-tier Anthropic keys hit per-minute token limits on the very first call because qmax-code's system prompt + tool catalog is large. The 429 was surfaced raw with no retry — blocking any trial user on a free key.

• Adds doWithRetry() in internal/agent/retry.go
• Retries up to 3 times on HTTP 429: 1 s → 2 s → 4 s backoff
• Parses Retry-After header to use the server-suggested delay when present
• Respects context cancellation during sleep (Ctrl+C exits immediately, doesn't hang for 4 s)
• Prints ● rate limited — retrying in 2s (1/3)… via the terminal status line
• Wired into both callStreamingAPI() and callAPI(); request body is rebuilt from the already-marshaled []byte slice each attempt (no double-read issue)

Closes QUA-645

Test plan

• With a free-tier key, trigger a large-context call — should see the retry status line and succeed on retry rather than dying immediately
• Verify Ctrl+C during a retry sleep exits cleanly rather than waiting out the delay
• Verify a non-429 error (e.g. 401, 500) is still surfaced immediately without retry

🤖 Generated with Claude Code

[qualitymaxapp]

✅ QualityMax Pipeline

Gate	Result
🔍 AI Review	✅ Clean
🧪 Repo Tests	✅ 298/298 passed (go)
🤖 AI Tests	✅ 28/32 passed

_{Powered by QualityMax — AI-Powered Test Automation}