split_large_files: RecursionError when a single parquet row exceeds the target size

Description

split_large_files.py's _split_table recurses forever (ends in RecursionError) when a parquet chunk contains a single row that is itself larger than the target size. There is no base case for num_rows <= 1, so the bisection keeps slicing the same 1-row table in half (1 // 2 = 0) forever.

This is realistic for interleaved-document parquet files with large embedded payloads (e.g. base64 images) — one 3 MB row with --target-size-mb 1 kills the tool.

Steps to reproduce

import pyarrow as pa
from nemo_curator.utils.split_large_files import _split_table

t = pa.table({"text": ["x" * 3_000_000]})
list(_split_table(t, 1024 * 1024))
# RecursionError

Expected behavior

A row that can't be split further becomes its own shard (same behavior the JSONL splitter documents for oversized single records — write it alone, ideally with a warning), and the tool finishes.

Actual behavior

_split_table slices [0 : n//2] and [n//2 : n]; for n=1 both slices are 1-row/0-row tables that never get smaller → infinite recursion.

Environment

nemo-curator main (93b48525), Python 3.12

Start in split_large_files.py at the _split_table entry point and run the provided pyarrow reproduction with a single 3 MB row and a 1 MB target. Ensure a row that cannot be split further is emitted as its own shard rather than recursing, and verify the reproduction completes without RecursionError.