Linear Loss Tracker

[pbontrager]

We recently rolled out linear losses which can further save memory over just chunked loss. It's been rolled out to all of the SFT recipes but there still remains followup work to implement it in a few more recipes and get it fully interoperable with compile + TP. These changes need to land before we can fully deprecate CEWithChunkedOutputLoss.

Original PRs:

• Refactor losses instantiation and chunked CE #2531
• Re-refactor Linear Loss #2667