docs: 修正文档错误并补充训练参数说明

- README: 补充训练参数速查表，完善训练命令示例
- design.md: 同步 inference 类图（SlotAllocator、GenerationParams、采样策略等
  新增类），修正参数名和类型错误，统一泛型符号
- params.md: 修正默认值（batch_size=1、num_workers=4），移除不存在参数
  （grpo_*、model_type、resume_dir），补充完整示例
- dataflow.md: _RadixNode 命名修正

assets/docs/dataflow.md

@@ -176,7 +176,7 @@ flowchart LR
 - **`TaskStatus`**: Task state enumeration
 - **`apply_sampling_strategies`**: Applies temperature, top-k, top-p sampling to logits
 - **`PrefixCacheManager`**: Radix tree-based prefix cache with LRU eviction for efficient KV cache reuse
-- **`RadixNode`**: Tree node structure for prefix caching
+- **`_RadixNode`**: Tree node structure for prefix caching
 - Continuous batching: new requests can join at any time, completed requests are released immediately
 
 #### 6.3 Server (`server.py`)