Fix silent data loss in push_to_hub when num_proc > num_shards - #8044

Summary

When num_proc exceeds the number of output shards in Dataset.push_to_hub, workers
without assigned output shards silently drop their data. The metadata still reports the
correct dataset length (using len(self[split])), masking the data loss.

This fix caps num_jobs at num_shards in _push_parquet_shards_to_hub, matching
what the streaming path already does.

Reproduction

1. Have a dataset with a small number of shards (e.g. 2)
2. Call push_to_hub with num_proc > num_shards (e.g. 6)
3. Only a fraction of the samples are actually written to the hub