Splits conflict between PEDES and SYSU

The train/test/val splits differ between the PEDES and SYSU datasets. Currently, for each SYSU split, we generate two files: train and test. This approach is problematic when we want to use our annotations to evaluate models on PEDES dataset.

To address this, we'll generate a single annotations.csv file. This file will introduce two new columns: split_pedes and split_sysu. The split_pedes column already exists. For split_sysu, we'll label training samples as "train" and either "query" or "gallery" for the respective samples.

Furthermore, we'll refactor parts of the code to accommodate this significant change. One such modification will be setting the distractor annotations bbox columns to NaN.