model load from recorder.load_object() missing _fitted variable

🐛 Bug Description

model loaded from recorder.load_object() missing _fitted variable. Not sure whether it is caused by pickle.dump or pickle.load

To Reproduce

Steps to reproduce the behavior:

1. model = init_instance_by_config(task["model"]), the model can be LSTM from pytorch

with R.start(experiment_name="train_model"):
    R.log_params(**flatten_dict(task))
    model.fit(dataset, save_path = ".model-pytorch")
    R.save_objects(trained_model=model)
    recorder = R.get_recorder()

with R.start(experiment_name="backtest_analysis"):
    recorder = R.get_recorder(rid, experiment_name="train_model")
    model = recorder.load_object("trained_model")

    # prediction
    recorder = R.get_recorder()
    ba_rid = recorder.id
    sr = SignalRecord(model, dataset, recorder)
    sr.generate()

It will return error saying _fitted not exist.

Expected Behavior

recorder.load_object should be able to load all variables in LSTM class

Environment

Note: User could run cd scripts && python collect_info.py all under project directory to get system information
and paste them here directly.

• Qlib version:0.6.1.dev0
• Python version: 3.7
• OS (Windows, Linux, MacOS): Linux
• Pytorch: 1.7.1+cu101

Start by reproducing the reported flow with init_instance_by_config, model.fit, R.save_objects, recorder.load_object, and SignalRecord.generate using the stated Python and PyTorch versions. Compare the LSTM model state before saving and after loading, focusing on the missing _fitted variable and whether pickle serialization or deserialization loses it. Done means the loaded model can be passed to SignalRecord.generate without the _fitted error.