Add rdma channel

[JimyMa]

⚠️ Breaking Changes
API Refactoring: Significant changes to the Python API. Previous methods like write_batch and read_batch have been replaced by unified write and read interfaces. Input arguments have shifted from object lists to vectorized parameters (e.g., vector<uintptr_t>) to reduce Python overhead.

Async Pattern Shift: The explicit async_op flag has been removed in favor of a full Future-based pattern. All communication operations (Send/Recv/Read/Write) now return Future objects (e.g., SlimeSendFuture), requiring the user to explicitly call .wait() for synchronization.

🚀 New Features & Core Improvements
Unified Endpoint Architecture:

Introduced the RDMAEndpoint class, which unifies the previously separated RDMAIOEndpoint (for one-sided RDMA) and RDMAMsgEndpoint (for two-sided messaging). Users can now manage multiple communication modes via a single endpoint instance.

Async Worker Runtime:

Implemented RDMAWorker and GlobalWorkerManager. A dedicated background thread now handles Completion Queue (CQ) events without relying on main-thread polling.

Optimized Resource Management:

Introduced GlobalContextManager and pooled RDMAContext management. This supports the reuse of contexts across multiple devices, reducing initialization overhead.

PyTorch Backend Upgrade:

Rewrote csrc/torch/slime_backend.cpp. The PyTorch distributed backend has been migrated to the new Unified Endpoint and Worker architecture, improving efficiency when used as a ProcessGroup.

Enhanced Developer Experience:

Added dlslime/_slime_c.pyi type stubs, significantly improving code completion and type checking in IDEs.

⚠️ 破坏性变更 (Breaking Changes)
API 重构：Python 侧的 API 发生了重大变化。旧有的 write_batch、read_batch 等方法被替换为统一的 write、read 接口，且参数传递方式由原来的对象列表变为向量化参数（如 vector<uintptr_t>），以减少 Python 层的开销。

异步模式变更：不再通过参数控制 async_op，而是全面转向 Future 模式。所有通信操作（Send/Recv/Read/Write）现在均返回 Future 对象（如 SlimeSendFuture），用户需显式调用 .wait() 来同步操作。

🚀 新特性与核心改进
统一端点架构 (Unified Endpoint)：

引入了全新的 RDMAEndpoint 类，将原先分散的 RDMAIOEndpoint（负责单边读写）和 RDMAMsgEndpoint（负责双边消息）进行了统一封装。用户现在可以通过一个端点实例同时管理多种通信模式。

异步 Worker 机制 (Async Runtime)：

新增 RDMAWorker 和 GlobalWorkerManager。引入了独立的后台线程来处理完成队列（CQ）事件，不再依赖主线程的轮询。

资源管理优化：

引入 GlobalContextManager 和 RDMAContext 池化管理，支持多设备上下文的复用，减少了资源初始化开销。

PyTorch 后端升级：

重写了 csrc/torch/slime_backend.cpp，将 PyTorch 分布式后端迁移至新的 Unified Endpoint 和 Worker 架构，提升了作为 ProcessGroup 使用时的效率。

开发体验提升：

新增 dlslime/_slime_c.pyi 类型桩文件（Type Stubs），显著改善了 IDE 中的代码补全和类型检查体验。