Fix: DeBERTa-v3 fp16 crash on T4 — use fp32, batch=2, grad_accum=16

Update notebook for T4 GPU: batch=4, grad_accum=8, eval_batch=8

Add ClauseGuard v4 Colab notebook: DeBERTa-v3-large 2-stage training (LEDGAR→CUAD) with ASL