fix(train): Fix step & destroy group

[xingchensong]

1. wenet 2.2.1 中 configs['step']从-1开始而不是0开始 https://github.com/wenet-e2e/wenet/blob/v2.2.1/wenet/bin/train.py#L223
2. wenet 2.2.1 中 self.step只有在accum_grad处才+1 https://github.com/wenet-e2e/wenet/blob/v2.2.1/wenet/utils/executor.py#L109 (上述不同应该不影响训练结果（或者说影响不大）。configs['step'] 对学习率影响只有一步之差。 self.step纯用来打印log，不影响训练流程)
3. dist.barrier() 移动到 new_group 之前，强制同步，减少超时出现的频率
4. 增加 destroy_process_group，对已经完成使命的group_join进行析构，避免 超过 ulimit -u 中限制的 max user processes RuntimeError: Resource temporarily unavailable due to running out of threads (ulimit -u) jax-ml/jax#2685

Root Cause (first observed failure):
[0]:
  time      : 2023-11-28_05:28:31
  host      : hobot-job-10017872-task-1.hobot-job-10017872.project-2080ti-speech-smalltask.svc.cluster.local.
  rank      : 0 (local_rank: 0)
  exitcode  : 1 (pid: 1554)
  error_file: /tmp/torchelastic_qjva9q5p/2023_fzr5792e/attempt_0/0/error.json
  traceback : Traceback (most recent call last):
    File "/usr/local/lib/python3.8/site-packages/torch/distributed/elastic/multiprocessing/errors/__init__.py", line 346, in wrapper
      return f(*args, **kwargs)
    File "wenet/bin/train.py", line 129, in main
      group_join = dist.new_group(backend="gloo",
    File "/usr/local/lib/python3.8/site-packages/torch/distributed/distributed_c10d.py", line 3335, in new_group
      pg = _new_process_group_helper(
    File "/usr/local/lib/python3.8/site-packages/torch/distributed/distributed_c10d.py", line 862, in _new_process_group_helper
      pg = ProcessGroupGloo(prefix_store, group_rank, group_size, timeout=timeout)
  RuntimeError: Resource temporarily unavailable

[xingchensong]

之前第140epoch 必触发 Resource temporarily unavailable, 现在 destroy_process_group 后，140ep后均可以继续训练，且趋势和wenet v2.2.1 一致

[xingchensong]

FYI: