[Feature Request] Move the model output bias to atomic model

[wanghan-iapcm]

Summary

current the bias is implemented as an attribute of the fitting network. however, not all atomic models has a fitting net.
conceptually it is better to implement the bias as an attribute of an atomic model.

Detailed Description

1. the bias is an attribute of an atomic model.
2. an atomic model implements the compute_output_stats method. the BaseAtomicModel may provide a default implementation.
3. the bias is automatically applied in the forward_common_atomic of the BaseAtomicModel.

Further Information, Files, and Links

No response