feat: ML模型波长记忆 + 跨传感器光谱重采样 + 数据加载规范化

训练端 (modeling_batch.py):
- _extract_train_wavelengths() 从列名提取训练波长写入 metadata
- load_data_batch/load_data_single 改为基于列名语义智能提取特征
- 废除 feature_start_column 硬编码位置索引

推理端 (inference_batch.py):
- preprocess_spectra() 分支A: train_wavelengths存在→np.interp精确重采样
- preprocess_spectra() 分支B: 无波长元数据→解析列名→np.interp到400-800nm标准网格
- 废除暴力截断和零值填充

src/core/steps/modeling_step.py

@@ -129,7 +129,7 @@ class ModelingStep:
 
     @staticmethod
     def train_models(
-        feature_start_column: str = "374.285004",
+        feature_start_column: Union[int, str, None] = None,
         preprocessing_methods: Optional[List[str]] = None,
         model_names: Optional[List[str]] = None,
         split_methods: Optional[List[str]] = None,