refactor: 重构打包模块，新增 BFD/FFD/Greedy 三种 bin-packing 算法，默认 BFD

- 将 pipeline/packing.py 拆分为 packing/ 子包 (base/stream/binpack)
- 新增 BfdPacker(默认)/FfDPacker/GreedyPacker，移除 StreamingPacker
- 超长序列直接截断至 pack_size
- group_size 语义改为"每 N 个 chunk 合并为一块"，默认 1000
- 新增 AutoTokenizer.token_to_id()，修复 ChatML 中 hacky 的 nl_id 获取
- pad_value 默认改为 2(pad_token_id)，position_ids pad=0, loss_mask pad=False
- 新增 position_ids 打包后归零一致性测试
- scripts/cache_h5.py 新增 --pack-algo 参数

pipeline/packing/base.py

@@ -0,0 +1,49 @@
+from abc import ABC, abstractmethod
+from typing import List, Optional, Union
+
+import torch
+from torch import Tensor
+
+
+class BasePacker(ABC):
+    """Abstract base class for sequence packing algorithms.
+
+    All packers must implement pack() and reset().
+    pack() takes a list of 1D tensors and returns a list of packed fixed-size tensors.
+    """
+
+    def __init__(
+        self,
+        pack_size: int,
+        pad_value: Union[int, bool] = 0,
+        dtype: Optional[torch.dtype] = None,
+    ):
+        self.pack_size = pack_size
+        self.pad_value = pad_value
+        self.dtype = dtype
+
+    @abstractmethod
+    def pack(self, sequences: List[Tensor]) -> List[Tensor]:
+        """Pack sequences into fixed-size chunks."""
+        ...
+
+    @abstractmethod
+    def reset(self) -> None:
+        """Reset packer state for instance reuse."""
+        ...
+
+    def _validate_and_normalize(self, sequences: List[Tensor]) -> List[Tensor]:
+        """Validate 1D tensors and unify dtype."""
+        if self.dtype is None and sequences:
+            self.dtype = sequences[0].dtype
+
+        normalized: List[Tensor] = []
+        for i, seq in enumerate(sequences):
+            if seq.dim() != 1:
+                raise ValueError(
+                    f"Expected 1D tensor at index {i}, got {seq.dim()}D tensor with shape {seq.shape}"
+                )
+            if seq.dtype != self.dtype:
+                seq = seq.to(self.dtype)
+            normalized.append(seq)
+        return normalized