fix(python): handle empty filtered shards in ShardedBatchSampler

[xushiyan]

Summary

• Fix ShardedBatchSampler._shard_scan so filtered scans with zero matches or empty per-rank shards yield an empty stream instead of raising.
• Preserve global round-robin sharding across filtered scanner batches by carrying rows_to_skip from the full batch length, not a sliced view.
• Skip batch.take([]) when a rank owns no rows in a batch (PyArrow has no array_take(int64, null) kernel).
• Add regression tests for empty shard, zero-match filter, and cross-fragment carryover.

Problem

When ShardedBatchSampler is called with a filter, the filtered path can crash in two cases:

1. Global zero-match — filter matches no rows -> pa.Table.from_batches([]) raises ValueError.
2. Empty per-rank shard — sparse filter with world_size > 1 leaves some ranks with zero rows -> batch.take([]) raises ArrowNotImplementedError.

The unfiltered path already handles emptiness; this aligns the filtered path with that contract.

Test plan

• test_sharded_batch_sampler_empty_filtered_shard (randomize on/off)
• test_sharded_batch_sampler_filtered_carryover_across_fragments
• uv run pytest python/tests/test_sampler.py (full sampler suite)

Found during distributed training work with sparse segment filters.