feat: 服务化基础设施 - 有界队列/超时/优雅关闭/metrics

- astrai/inference/scheduler.py: 有界队列 (max_queue_size) 拒绝满时入队抛 RuntimeError
    -> 请求超时检测 (deadline + _abort_expired_tasks)，超时任务 abort 释放页并通知回调
    -> stop() 改为 drain 模式：等待活跃任务自然结束再强制清理
    -> get_stats() 扩展 latency P50/P95/P99 + cache hit rate
- astrai/inference/engine.py: generate/generate_async 新增 timeout 参数
    -> _generate_streaming/_generate_non_streaming 捕获 add_task 异常并清理
- astrai/inference/server.py: 新增 /metrics 端点 (Prometheus 格式)
    -> chat completions 端点捕获 RuntimeError 返回 503
    -> configure_server 传递 max_queue_size/request_timeout
- astrai/inference/cache.py: 新增 lookup_hits/lookup_misses 计数器
- tests/: fix stats key total_tasks -> total_requests

tests/inference/test_scheduler_concurrency.py

@@ -173,5 +173,5 @@ def test_scheduler_concurrent_get_stats(mock_model_and_tokenizer):
 
     # Verify stats are consistent
     for stats in results["stats"]:
-        assert "total_tasks" in stats
-        assert stats["total_tasks"] >= 0
+        assert "total_requests" in stats
+        assert stats["total_requests"] >= 0