[LLaVA-OV] Support LLaVA-OneVision eval loss and data filtering

[Caozhou1995]

This PR adds three important features:

• In the eval mode, the loss corresponding to each sample is output, so that data filtering can be performed subsequently. To enable this feature, simply set skip_train: True, and the rest of the settings are consistent with those in the train mode.
• Data filtering is provided, which allows custom filtering rules, reads the loss corresponding to the sample from the log, and stores the sample id that need to be retained in a JSON file.Usage: python filter_to_json.py --input_dir outputs --output sample_ids.json
• When processing data, the input of filter_json is added, and if the sample ID is not in this JSON file, it can be skipped. Usgae: bash make_llava_ov_wds.sh $DATA_PATH $EXPNAME_PATH $HOSTFILE $FILTER_JSON
• If you do not want to make it from scratch and only filter it from the already made tar file according to filter_json, Usage: python filter_by_json.py --input_dir raw_dir --output_dir output_dir --json_file filter_json，and you can use mpirun to start multiple nodes to process.