Consolidate all training logging through MetricsLogger

Log patience counter, best val loss/step in val records

Per-model early stopping: freeze converged variants individually

Revert batch size default to 256

Push metrics to HF at eval intervals, add dashboard HF sync

Add early stopping patience to multi-model training

Remove .item() CUDA sync from hot path, batch size 512, run slugs

Add post-training evals, /dev/shm checkpoints, async HF push, and _orig_mod fix

Safetensors migration, checkpoint integrity, and multi-model training. (#1)