When batch size is bigger than dataset size, the SFT finishes without error, but not doing any grad updates

**Describe the bug**

E.g. if bs=1024, but dataset only has 128 samples, I just see a bunch of these in the logs

========================= Epoch 1/2000 =========================

========================= Epoch 2/2000 =========================

========================= Epoch 3/2000 =========================

========================= Epoch 4/2000 =========================

========================= Epoch 5/2000 =========================

========================= Epoch 6/2000 =========================

========================= Epoch 7/2000 =========================

========================= Epoch 8/2000 =========================

and then the final checkpoint is saved but is the same as original.

**Steps/Code to reproduce bug**

Run any SFT with batch size bigger than dataset size

**Expected behavior**

Either fail with error or reduce batch size automatically to be equal to the dataset size.