feat(map): reuse unchanged columns when input_columns specified to reduce disk usage (#6013)

[ArjunJagdale]

Summary

This PR addresses #6013 by reusing unchanged columns from the original dataset in the map() method when input_columns is specified.

What’s Implemented

• Injected logic at the end of Dataset.map() to:
  • Identify untouched columns not in input_columns or remove_columns
  • Select those columns from the original dataset
  • Concatenate them with the transformed result using pyarrow.concat_tables

Example Behavior

ds = Dataset.from_dict({"a": [1, 2], "b": [3, 4]})
ds2 = ds.map(lambda x: {"c": x["a"] + 10}, input_columns=["a"], remove_columns=["a"])
print(ds2.column_names)  # Output: ['b', 'c']

Column b is reused from the original dataset.

Notes

• This keeps disk usage and caching minimal by avoiding full dataset duplication.
• Only triggered when input_columns is set.

---

cc @lhoestq @mariosasko for review 🙂