[BUG] DataFrame.to_arrow is inconsistent with pa.Table.from_pandas() when preserve_index=True

[rjzamora]

Describe the bug
It is my understanding that we want the DataFrame.to_arrow API to be consistent with pa.Table.from_pandas() (when possible). This is currently not the case when DataFrame.index is a RangeIndex, and preserve_index=True is specified. In this case, pa.Table.from_pandas() will use the RangeIndex information to produce an explicit "__index_level_0__" column in the output pyarrow Table.

Side Note: Creating an explicit column is the only way to generate a Table schema that will "safely" preserve the index in dask, because that schema may be used to "rebuild" partitions with a different number of rows later on (and so the original start/stop metadata can be "wrong").

Steps/Code to reproduce bug

In [1]: import pandas as pd, cudf
In [2]: import pyarrow as pa
In [3]: df = cudf.DataFrame({"x": ["cat", "dog"] * 5})
In [4]: pdf = df.to_pandas()

The schema extracted from pandas will have an "__index_level_0__" column when preserve_index=True, and will store RangeIndex metadata if preserve_index=True or preserve_index=None:

In [6]: pa.Table.from_pandas(pdf, preserve_index=None).schema
Out[6]: 
x: string
-- schema metadata --
pandas: '{"index_columns": [{"kind": "range", "name": null, "start": 0, "' + 362
In [7]: pa.Table.from_pandas(pdf, preserve_index=True).schema
Out[7]: 
x: string
__index_level_0__: int64
-- schema metadata --
pandas: '{"index_columns": ["__index_level_0__"], "column_indexes": [{"na' + 431

The schema extracted from cudf will never have an explicit index column, and will only store RangeIndex metadata if preserve_index=True.

In [8]: df.to_arrow(preserve_index=None).schema
Out[8]: 
x: string
-- schema metadata --
pandas: '{"index_columns": [], "column_indexes": [], "columns": [{"name":' + 177
In [9]: df.to_arrow(preserve_index=True).schema
Out[9]: 
x: string
-- schema metadata --
pandas: '{"index_columns": [{"kind": "range", "name": null, "start": 0, "' + 362

Expected behavior
I'd like for DataFrame.to_arrow(preserve_index=True) to be consistent with pa.Table.from_pandas(..., preserve_index=True). More specifically, I'd like cudf to produce an explicit column in the pyarrow Table. When the original RangeIndex is un-named, it probably makes sense to call the column "__index_level_0__". However, DataFrame.from_arrow will also need to recognize that "__index_level_0__" should round-trip to an un-named index.

Additional context
This bug complicates #13893, because distributed's "p2p" shuffle now assumes that to_pyarrow_table_dispatch will include the actual data for an index column when preserve_index=True (not just range metadata). We can certainly include a band-aid for this in dask_cudf, but the best long-term fix belongs in cudf.