Improve uv and flash-attn installation

[shiweijiezero]

Summary

• add scripts/install/install_trinity.sh to provide a one-command uv setup for the documented vLLM environment
• add uv overrides for two known dependency conflicts:
  • use numpy>=2 because vLLM's OpenCV dependency requires NumPy 2+, while veRL declares numpy<2
  • use transformers>=5.12.1 because SGLang 0.5.13 pins Transformers 5.8.1, while Trinity requires Transformers 5.12.1 or later
• update the flash-attn helper because the previous implementation was tied to flash-attn 2.8.1 and a CUDA 12 OSS URL scheme; the new implementation discovers a compatible wheel using the installed Python, PyTorch, CUDA, and platform
• require vLLM 0.22 or later because Trinity uses the pack_tensors and unpack_tensor weight-transfer APIs introduced in vLLM 0.22
• align the README and installation documentation with pyproject.toml, which requires flash-attn>=2.8.3, document the scripted and manual uv installation flows, and clarify that .venv should be activated before invoking trinity directly