Full-parameter SFT launcher enables LoRA instead of full fine-tuning

### Affected revision

`main` at `3d1965cac5529b72c9dab9af7a3f456c5bbd7172`.

### Reproduction

1. Inspect `UniVR_SFT/scripts/train_sft_full.sh`: it passes `--use_lora True`.
2. Inspect `UniVR_SFT/train.py`: when `use_lora` is true, the model is wrapped with `get_peft_model` and only LoRA adapters are trained.
3. Compare `train_sft_lora.sh`, which intentionally passes the same value.

### Expected behavior

The launcher named `train_sft_full.sh` should select full-parameter fine-tuning, consistent with the `LoraArguments.use_lora` help text (`False` means full fine-tuning).

### Actual behavior

The full-parameter launcher enables LoRA, so invoking it performs parameter-efficient adapter training instead of updating the full model.

### Impact

Training runs started through the advertised full-parameter script use a different optimization mode and produce different checkpoint contents than the script name promises. This can silently invalidate expensive experiments.

### Suggested fix

Set `--use_lora False` in the full-parameter launcher and add a small regression check that keeps the full and LoRA launchers distinct.

Start with UniVR_SFT/scripts/train_sft_full.sh and compare its use_lora argument with UniVR_SFT/scripts/train_sft_lora.sh. Read the use_lora handling in UniVR_SFT/train.py and add the suggested regression check. Done means the full launcher selects full-parameter fine-tuning while the LoRA launcher remains distinct.

Written by the indexing model from the issue text.