add process sync in evaluation metric computation via a temp file in lmms_eval/evaluator.py

[Dousia]

While evaluating model responses with metrics that require GPUs to perform computation, the current codebase uses cuda:0 to compute the metric score and lets other GPU processes waiting in barrier. The barriers take up a large amount of computation, slowing down metric computation on these GPUs. For metrics like CAPTURE used for detail image caption evaluation, this leads to intolerably long evaluation time.
In this PR, a process synchronization operation is added to metric evaluation code. This operation synchronizes processes with a temp file stored in {cli_args.output_path}/rank0_metric_eval_done.txt. This synchronization does not take up the computation of GPUs with rank>0.