[Feature Request] Use shared_dict for linear model

[ChiahsinChu]

Summary

Currently, the linear model is implemented to combine multiple frozen models. However, this architecture can be further used for models with more complicated combining rules. I wonder if it is possible to have the design similar to multi-task/model/shared_dict and make all submodels share (some of) their parameters.
I would like to know the comments/suggestions from developers about this feature and the relevant code implementation. I can implement it once we come to a conclusion.
Thanks.

Detailed Description

An example input expected:

"model": {
    "type": "custom_ener_model",
    "shared_dict": {
      "type_map_all": [
        "O",
        "H"
      ],
      "sea_descriptor_1": {
        "type": "se_e2_a",
        "sel": [
          46,
          92
        ],
        "rcut_smth": 0.50,
        "rcut": 6.00,
        "neuron": [
          25,
          50,
          100
        ],
        "resnet_dt": false,
        "axis_neuron": 16,
        "type_one_side": true,
        "seed": 1,
        "_comment": " that's all"
      },
      "_comment": "that's all"
    },
    "models": [
    {
        "type_map": "type_map_all",
        "descriptor": "sea_descriptor_1",
        "fitting_net" : {
	    "neuron":		[240, 240, 240],
	    "resnet_dt":	true,
	    "seed":		1
	}
    },
    {
        "type_map": "type_map_all",
        "descriptor": "sea_descriptor_1",
        "fitting_net" : {
            "type": "dipole",
	    "neuron":		[100, 100, 100],
	    "resnet_dt":	true,
	    "seed":		1
	}
    }
    ],
    "model_arg":{}
},

Further Information, Files, and Links

No response

[iProzd]

@ChiahsinChu LGTM to add a 'share_dict' param inside the linear_ener model. The code for preprocess input and sharing params can be found in https://github.com/deepmodeling/deepmd-kit/blob/devel/deepmd/pt/utils/multi_task.py#L14 and https://github.com/deepmodeling/deepmd-kit/blob/devel/deepmd/pt/train/wrapper.py#L62, respectively.