fix(agents): generalize rate-limit retry for all LLM providers

Replace OpenAI-specific RateLimitError catch with a generic
_is_rate_limit_error() helper that detects rate limits from:
- OpenAI (RateLimitError)
- Anthropic (status_code=429)
- Google (ResourceExhausted / 429)
- Any provider with 'rate limit' in error message

Fixes #3882

Author: guoyangzhen