Feature request: built-in loss extrapolator

### 🚀 The feature, motivation and pitch

According to the Neural Scaling Laws paper (Kaplan, et. al. 2020) we can make a pretty good guess at what a model's loss will be after just a few training steps. It would be nice to include this in the training printout as like a "estimated final loss" or something.

### Alternatives

_No response_

### Additional context

_No response_

The issue names only the training printout and does not identify a file, test, estimator interface, or output format. Start by locating the code that produces training loss output and review how training metrics are reported. Before implementation, clarify the scaling-law method, required inputs, display format, and tests that would define the estimate as done.

Written by the indexing model from the issue text.