Vbench

Bug fixes
- Fix OOM error in `aesthetic_quality` dimension (1a213e4fd100025a58ecb7f08537202f0a3aef28)

Features
- Support inference with multi-gpu using `torchrun`, see usage in README.md (fc4620b4d620114b77842c336f7a193c37e8e12e)
- `vbench evaluate` command to support launching multi-gpu inference (e3e66641826978192923ad19afefc78132853a36)

Bug fixes
- Fix preprocessing for high-resolution images/videos in imaging_quality (46275d9, b484867)
- Fix VBench_full_info.json not found in PyPI package (90f0973)

Features
- Add `filter_scope` argument for selecting the videos to be considered for filtering out static videos (fb75b02)
- Add `mode` argument to select between three different evaluation modes (02e2ee1):
> custom_input: Receive input prompt from --prompt/--prompt_file flags or read directly from the filename
> vbench_standard: Evaluate using the standard prompt suite of VBench (default)
> vbench_category: Evaluate using a specific category of VBench
- Allow evaluating with prompts specified by a file containing a dictionary using `--prompt_file`. (efaa6d3)

Others Changes
- Receive list of dimension instead of single dimension in `--dimension` flag (91bece8)

This release contains:

- Pretrained model weights for all 16 T2V evaluation dimensions
- VBench prompt suite for T2V
- Source code for evaluating all 16 T2V dimensions
- `vbench evaluate` command to evaluate T2V on vbench prompt suite (cc68b18)
- `static_filter` command for filtering out static videos (for temporal flickering dimension) (fd3613a)