[Feature Request] Implement `out_bias` and `out_std` in the TensorFlow backend

[njzjz]

Summary

Implement out_bias and out_std of a model in the TensorFlow backend.

Detailed Description

Currently, out_bias and out_std of a model are implemented in dpmodel, PyTorch, Paddle, but not in TensorFlow. TensorFlow only implements Fitting-level's bias_atom_e, which is a bit different than out_bias - the shape of bias_atom_e is embedding_width while the shape of out_bias is dim_out, which are the same for ener, dos, and property, but different for dipole and polar.

There is also a workaround to convert between out_bias and bias_atom_e, which causes some problems in dipole. After implementing out_bias, this workaround could be cleaned.

The implementation should follow other variables, which make the variable serializable, deserializable, and restorable.

Further Information, Files, and Links

No response