Deps: Pin requirements for reproducible installs (Torch 2.3.1 cu121)

[Mars122]

Title: Pin requirements for reproducible installs (tested with Torch 2.3.1 cu121)

Summary

This PR updates settings/requirements.txt to include specific versions for common dependencies, improving reproducibility across machines/pods while aligning with the existing Torch/XFormers setup in settings/setup.sh.

What changed

• Pinned versions for frequently changing libraries: Pillow, imageio, imageio-ffmpeg, ipython, matplotlib, lpips, einops, safetensors, sentencepiece, accelerate, transformers, omegaconf, pyarrow, redis, loguru, pandas.
• Kept diffusers==0.32 as-is (already pinned). Torch/XFormers remain pinned in settings/setup.sh.
• Left kiui, image-reward, and deepspeed unpinned to avoid unexpected compatibility constraints, as their usable versions can vary by environment and training needs.

Rationale

• The project already guides users toward Torch 2.3.1 + cu121 and xformers 0.0.27; pinning the surrounding Python stack reduces churn and “works on my machine” divergences.
• The chosen versions are widely used and compatible with diffusers 0.32 and transformers 4.45.

Notes

• If desired, we can further constrain or relax specific pins; this PR aims for a conservative, reviewable baseline.
• No changes to core code or training/inference scripts.

Testing

• Verified a clean install on CUDA 12.1 images with Torch 2.3.1/cu121 and xformers 0.0.27.

License

• Requirements-only change.