VRTrainDataset ignores the cfg_prob constructor argument

### Affected revision

`main` at `3d1965cac5529b72c9dab9af7a3f456c5bbd7172`.

### Reproduction

Construct `VRTrainDataset` with any nonzero classifier-free-guidance probability:

```python
dataset = VRTrainDataset(dataset_cfg={}, tokenizer=tokenizer, cfg_prob=0.75)
print(dataset.cfg_prob)
```

The constructor accepts `cfg_prob`, but `UniVR_SFT/src/utils/dataset_vr_train.py` assigns `self.cfg_prob = 0.0` unconditionally. A focused run with `cfg_prob=0.75` prints `0.0`.

### Expected behavior

`dataset.cfg_prob` should preserve the caller-provided probability, so `_build_sample` can select the unconditional prompt at the configured rate.

### Actual behavior

Every instance stores `0.0`; therefore the later condition `self.cfg_prob > 0` is always false and classifier-free-guidance training cannot be enabled through this constructor argument.

### Impact

Callers can configure a nonzero probability without an error, but the requested training behavior is silently disabled for every sample.

### Suggested fix

Assign the constructor argument to `self.cfg_prob` and add a focused constructor regression test.

Start in UniVR_SFT/src/utils/dataset_vr_train.py at the VRTrainDataset constructor and inspect how cfg_prob is stored and later used by _build_sample. Add a focused constructor regression test using a nonzero probability, then verify the instance preserves the caller-provided value and the test passes.

Written by the indexing model from the issue text.