[Feature] Task retry and error recovery

## Summary Add configurable retry logic and error recovery strategies for failed tasks. ## Motivation In production environments, LLM API calls can fail due to rate limits, timeouts, or transient errors. Currently, a failed task triggers `cascadeFailure()` which marks all downstream tasks as failed. This is correct but aggressive — many failures are recoverable. ## Proposed Approach - Add `retryPolicy` to task configuration: ```typescript { maxRetries: 3, backoff: 'exponential', // or 'linear', 'fixed' retryableErrors: ['rate_limit', 'timeout'], } ``` - Retry at the task level (re-run the agent with the same prompt) - Only cascade failure after all retries are exhausted - Emit retry events for observability ## Acceptance Criteria - [ ] Configurable retry count and backoff strategy - [ ] Distinguish retryable vs non-retryable errors - [ ] Retry events emitted for monitoring - [ ] Tests for retry and eventual failure scenarios