fix(deps): use a floor pin for safetensors so fresh installs resolve

[rakhimovv]

Fixes #8.

requirements.txt is installed after pip install diffsynth==2.0.11, so its exact pin safetensors==0.7.0 downgrades the version that the diffsynth dependency chain had just resolved. diffsynth declares transformers with no upper bound, and transformers raised its floor to safetensors>=0.8.0 in 5.13.0 (2026-07-03), enforced at import time. The result is that a fresh install following the README fails on import diffsynth:

ImportError: safetensors>=0.8.0 is required for a normal functioning of this module, but found safetensors==0.7.0.

This changes the exact pin to a floor, which lets pip satisfy both constraints and keeps working if transformers raises its floor again.

Safe for this codebase: safetensors 0.8.0's breaking change is limited to the low-level serialize / serialize_file API, while this repo and diffsynth only use the read path (from safetensors import safe_open).

Verified on Linux / 4xH100 / CUDA 12.8, Python 3.10.21, torch 2.8.0+cu128:

• pip install resolves cleanly (safetensors-0.8.0, transformers-5.15.0); pip check reports no broken requirements.
• bash scripts/infer_example.sh runs to completion on all three demo/demo.jsonl episodes, producing outputs/inference/episode{0,1,2}.mp4.