Parquet uploads off-by-one naming scheme

### Describe the bug

I noticed this numbering scheme not matching up in a different project and wanted to raise it as an issue for discussion, what is the actual proper way to have these stored?

The `-SSSSS-of-NNNNN` seems to be used widely across the codebase. The section that creates the part in my screenshot is here https://github.com/huggingface/datasets/blob/main/src/datasets/arrow_dataset.py#L5287
There are also some edits to this section in the single commit branch.

### Steps to reproduce the bug

1. Upload a dataset that requires at least two parquet files in it
2. Observe the naming scheme

### Expected behavior

The couple options here are of course **1. keeping it as is**

**2. Starting the index at 1:**
train-00001-of-00002-{hash}.parquet
train-00002-of-00002-{hash}.parquet

**3. My preferred option** (which would solve my specific issue), dropping the total entirely:
train-00000-{hash}.parquet
train-00001-{hash}.parquet

This also solves an issue that will occur with an `append` variable for `push_to_hub` (see https://github.com/huggingface/datasets/issues/6290) where as you add a new parquet file, you need to rename everything in the repo as well.

However, I know there are parts of the repo that use 0 as the starting file or may require the total, so raising the question for discussion.

### Environment info

- `datasets` version: 2.14.6.dev0
- Platform: macOS-14.0-arm64-arm-64bit
- Python version: 3.10.12
- Huggingface_hub version: 0.18.0
- PyArrow version: 12.0.1
- Pandas version: 1.5.3