Fix: unwrap DDP before enabling gradient checkpointing for HF compatibility

[magicwang1111]

This PR introduces three key improvements:

1. Dependency Standardization

   • Added a requirements.txt file based on the original uv.lock to ensure reproducible installations across environments.
   • Locks versions for diffusers, accelerate, transformers, CUDA toolkits, and other core libraries.

2. Support for Custom Master Port

   • Added a new --main_process_port option in scripts/train_distributed.py to allow explicit control over the master port used by Accelerate’s distributed launcher.
   • Prevents port conflicts when launching multiple distributed training jobs on the same host.

   python scripts/train_distributed.py
   configs/your_config.yaml
   --num_processes 2
   --main_process_port 29600

3. **Fix for HF DDP Compatibility

In src/ltxv_trainer/trainer.py, unwrap the base model from DistributedDataParallel before calling the gradient-checkpointing API.

Prevents the runtime AttributeError: 'DistributedDataParallel' object has no attribute 'enable_gradient_checkpointing'.