Send map workers only their shard's rows instead of the whole table

[behroozazarkhalili]

Resolves #1992

Root cause

Dataset.map(num_proc>1) shards the table with Arrow slices (table.slice), and Arrow slices share their parent buffers. When a shard is pickled for a worker, InMemoryTable.__getstate__ serializes the sliced arrays' full buffers, so every worker receives the whole table. A 2-process map therefore moved 2x the dataset through the pool queues, which is why it came out slower than a single process on in-memory datasets.

Fix

Right before each job is submitted to the pool, compact its shard to only its rows:

• Take the shard's indices when it carries an indices mapping (after select, filter, shuffle, train_test_split), so index-mapped shards no longer ship the full table either.
• Copy each chunk individually so the chunk layout is preserved (chunks above the 32-bit offset limit stay separate).
• Rebuild string_view / binary_view columns (top-level or nested) from their values, because concatenating a view array keeps the parent value buffer.
• In-memory ConcatenationTable inputs (the result of concatenate_datasets) are compacted through their combined table; memory-mapped tables are left alone because they pickle by path.
• Compaction happens per job, lazily, on a shallow copy of the shard, so fingerprints, features, format and cache-file names are unchanged, and the source shards are not mutated.
• Any Arrow error during compaction falls back to the uncompacted shard, so unsupported types still map exactly as before.

Measurements

100 unique 140-byte rows, num_proc=2, bytes of each worker's pickled shard and how many of the 100 source rows appear in it:

dataset	before	after
Dataset.from_dict	15,910 B, 100 rows	8,507 B, 50 rows
after non-contiguous select	16,967 B, 100 rows	8,507 B, 50 rows
concatenate_datasets (70 + 30 rows)	100 rows in one worker	50 rows
string_view column	100 rows	50 rows

Map results, order, features and fingerprints match upstream for all of these, including formatted (numpy/torch) datasets, Array2D extension columns, large types, and num_proc > num_shards.

Tests

New tests/test_map_multiprocessing.py: captures the exact bytes each worker receives and asserts other shards' rows are absent for plain, indexed, concatenated and view-typed inputs; checks results and fingerprints across memory/indexed/disk storage; preserves chunk boundaries above the offset limit and nested chunked features; and exercises the fallback path by injecting Arrow errors.