fix(arrow_dataset): initialise writer on first non-None map result, not on i==0

[s-zx]

Summary

Dataset.map crashes with AttributeError: 'NoneType' object has no attribute 'write' when the map function returns None for the first N examples and a dict for later ones (issue #7990).

Root Cause

The ArrowWriter was initialised with a guard tied to the example index:

# Non-batched path (line ~3736)
if i == 0:
    buf_writer, writer, tmp_file = init_buffer_and_writer()

# Batched path (line ~3762)
if i and i[0] == 0:
    buf_writer, writer, tmp_file = init_buffer_and_writer()

update_data is set lazily via prepare_outputs(): when the first example returns None, update_data = False and no writer is created. When example i=2 returns a dict, update_data flips to True, but since i != 0 the writer initialisation is skipped. The subsequent writer.write(example) call then crashes because writer is None.

Fix

Replace the index-based guards with if writer is None:

# Non-batched path
if writer is None:
    buf_writer, writer, tmp_file = init_buffer_and_writer()

# Batched path
if writer is None:
    buf_writer, writer, tmp_file = init_buffer_and_writer()

The writer is now created on the first example/batch that actually produces output, regardless of its position.

Tests

Added test_map_writer_initialized_when_first_examples_return_none to tests/test_arrow_dataset.py: applies a map function that returns None for indices 0–1 and a transformed dict for the rest, verifying the call completes without error and produces correct output.

Fixes #7990