Packing in DPOTrainer

### Feature request

packing can be supported in dpo trainer.

### Motivation

sequences with high variation in length padded together will waste a lot of resources. issue #1274 mentioned this, but i don't think the conclusion is correct. packing won't conflict with pairwise datasets. you just need to unpack the sequence after forwarding. you can easily identify the start and the end of each sequence by the position of 0 in `position_ids`.

### Your contribution

actually i already have a version of dpo trainer that can deal with packing:

```python
def _forward_one(self, model: nn.Module, batch: dict, name: str):
batch = dict(batch)
targets = batch.pop("targets")
loss_masks = batch.pop("loss_masks").to(torch.bool)
out = model(**batch)
logits = out["logits"]
targets = targets.clone()
targets[~loss_masks] = 0 # dummy token; we'll ignore the losses on these tokens later
per_token_logps = torch.gather(logits.log_softmax(-1), dim=2, index=targets.unsqueeze(2)).squeeze(2)
per_token_logps[~loss_masks] = 0
# unpack logp
assert batch["position_ids"].shape[0] == 1
starts = (batch["position_ids"] == 0).nonzero()[:, 1]
seqs = []
for i in range(len(starts) - 1):
seqs.append(per_token_logps[0, starts[i]: starts[i + 1]].sum())
seqs.append(per_token_logps[0, starts[-1]:].sum())
all_logps = torch.stack(seqs)
return {
f"{name}_logps": all_logps,
f"mean_{name}_logits": logits[loss_masks].mean(),
}

def concatenated_forward(self, model: nn.Module, batch: dict):
ret = {}
for name in ("chosen", "rejected"):
ret = {**ret, **self._forward_one(model, batch[name], name)}
return ret
```

this does not directly fit into the complex logic of dpo trainer, but the idea is that it is possible to do packing.