fix: correct discriminative LR and early-exit training

[JulesBelveze]

Summary

This fixes optimizer grouping across transformer architectures and corrects how DeeBERT and BERxiT train their exits.

• infer learning-rate schedules independently for each encoder and decoder stack
• assign embeddings a lower rate, task heads the top rate, and keep normalization weights out of weight decay
• follow DeeBERT's two-stage setup and BERxiT's alternating classifier and LTE objectives
• use standard ReduceLROnPlateau with a real epoch-loss monitor

The old optimizer-state remapping has been removed. Existing optimizer and scheduler state from the previous group layout needs to be restarted. Strict DeeBERT or BERxiT checkpoints from the old classifier and gate layout need a non-strict load or migration.

Validation

• 76 deterministic tests
• black, isort, flake8, scoped mypy, and diff checks