fix(pt): invalid type_map when multitask training (deepmodeling#4031)

It seems that in 3.0.0b3, executing multitask training or finetune task
would run into a `RuntimeError`, calling inconsistent type map. The
error log is shown below. However, the `type_map` in mutitask should be
a shared dict. Diving into the source code, we would see a `type_map`
[here](https://github.com/deepmodeling/deepmd-kit/blob/0e0fc1a63e478d3e56285b520b34a9c58488d659/deepmd/pt/entrypoints/main.py#L300).
It would cause an empty type_map in multitask training because of no
`type_map` found.

After applying the modification in this PR, everything seems to be well.

```
 Traceback (most recent call last):
    File "/public/home/ypliucat/.conda/envs/deepmd-kit-3.0.0b3/lib/python3.10/site-packages/torch/distributed/elastic/multiprocessing/errors/__init__.py", line 346, in wrapper
      return f(*args, **kwargs)
    File "/public/home/ypliucat/.conda/envs/deepmd-kit-3.0.0b3/lib/python3.10/site-packages/deepmd/pt/entrypoints/main.py", line 562, in main
      train(FLAGS)
    File "/public/home/ypliucat/.conda/envs/deepmd-kit-3.0.0b3/lib/python3.10/site-packages/deepmd/pt/entrypoints/main.py", line 311, in train
      train_data = get_data(
    File "/public/home/ypliucat/.conda/envs/deepmd-kit-3.0.0b3/lib/python3.10/site-packages/deepmd/utils/data_system.py", line 802, in get_data
      data = DeepmdDataSystem(
    File "/public/home/ypliucat/.conda/envs/deepmd-kit-3.0.0b3/lib/python3.10/site-packages/deepmd/utils/data_system.py", line 184, in __init__
      self.type_map = self._check_type_map_consistency(type_map_list)
    File "/public/home/ypliucat/.conda/envs/deepmd-kit-3.0.0b3/lib/python3.10/site-packages/deepmd/utils/data_system.py", line 616, in _check_type_map_consistency
      raise RuntimeError(f"inconsistent type map: {ret!s} {ii!s}")
  RuntimeError: inconsistent type map: ['Ag', 'Cu'] ['Ag', 'Ni']
```

<!-- This is an auto-generated comment: release notes by coderabbit.ai
-->

## Summary by CodeRabbit

- **New Features**
- Enhanced the training process to ensure consistent handling of model
type configurations, improving clarity and availability based on
multi-task settings.

<!-- end of auto-generated comment: release notes by coderabbit.ai -->

Signed-off-by: Futaki Haduki <812556867@qq.com>

deepmd/pt/entrypoints/main.py

@@ -297,8 +297,8 @@ def train(FLAGS):
             "Calculate neighbor statistics... (add --skip-neighbor-stat to skip this step)"
         )
 
-        type_map = config["model"].get("type_map")
         if not multi_task:
+            type_map = config["model"].get("type_map")
             train_data = get_data(
                 config["training"]["training_data"], 0, type_map, None
             )
@@ -308,6 +308,7 @@ def train(FLAGS):
         else:
             min_nbor_dist = {}
             for model_item in config["model"]["model_dict"]:
+                type_map = config["model"]["model_dict"][model_item].get("type_map")
                 train_data = get_data(
                     config["training"]["data_dict"][model_item]["training_data"],
                     0,