feat: support SFT directly from JSONL without dataset_config.json

- JsonlStore falls back to built-in messages config when no config file found and tokenizer_path is provided
- DatasetFactory.load forwards tokenizer_path to store for SFT/SEQ+jsonl
- assistant turns train, other roles masked, position_ids doc_reset

astrai/dataset/dataset.py

@@ -346,7 +346,15 @@ class DatasetFactory(BaseFactory["BaseDataset"]):
         if processor is not None:
             store.load(load_path, processor=processor, **kwargs)
         else:
-            store.load(load_path, **kwargs)
+            load_kwargs = dict(kwargs)
+            if (
+                tokenizer_path is not None
+                and storage_type == "jsonl"
+                and train_type in ("seq", "sft")
+                and "tokenizer_path" not in load_kwargs
+            ):
+                load_kwargs["tokenizer_path"] = tokenizer_path
+            store.load(load_path, **load_kwargs)
 
         return cls.create(train_type, store=store)