[Feature Request] Allow one model with multiple fitting networks

### Summary

Allow one model with multiple fitting networks.

### Detailed Description

In some energy models, there are multiple parameters dependent on the local environment. For example, one might want to predict the 1) atomic charges for the long-range electrostatic energy and 2) the short-range atomic energies in one energy model (i.e., 3rd-HDNNP proposed by J. Behler: 10.1103/PhysRevB.83.153101). I would expect the architecture to be somewhat similar to the multi-task training but with only one loss function.

### Further Information, Files, and Links

_No response_