[Epic] Instructlab "training" code consolidation for different GPU scenarios

[JamesKunstle]

Currently, training loops exist both in instructlab/instructlab and in this repo. Bringing all training-related paths under one repo would be helpful, especially regarding testing.

This would include:

1. CPU-only
2. MPS

3. =1 GPU (from the minimum-viable card up to 8xMI300x/H100)

The implication of this change might be that this library becomes a "functional-programming" alternative to the current typical pattern of a Trainer class, opinionated by our optimizations.

[JamesKunstle]

Related: #352