[python][torch] Support shuffle for Arrow and tensor batch streaming

[TheR1sing3un]

Purpose

Torch streaming currently rejects shuffle=True with batch_format="pyarrow" or "torch". Support bounded row shuffling for both batch formats so training can mix samples across reader batches while keeping payloads in Arrow until tensor conversion.

Reuse the existing seed/epoch and rank/worker semantics. Incoming Arrow blocks replace random slots in a rolling buffer, and bounded split interleaving mixes input sources. Filters and sharding precede shuffle; a binding limit retains the existing ordered selection. Output batching and custom tensor conversion run afterward. Close all active readers on completion, early termination, and read/conversion failures.

The default unshuffled path and existing row shuffle order are preserved. Document that this is worker-local buffer shuffling, with additional memory for incoming blocks, gathered output and format readers, and that prefetch_concurrency=1 still applies. Related to #9365.

Tests

• Torch suite: 69 tests passed initially; the existing standalone torchrun test hit this machine's hostname-resolution timeout, also reproduced on unchanged master. It passed when rerun with PET_LOCAL_ADDR=127.0.0.1 GLOO_SOCKET_IFNAME=lo0.
• Arrow 16 / NumPy 1.24 compatibility: 20 focused tests passed. Main environment: Python 3.11, Torch 2.8, Arrow 19.
• Covers Arrow/Tensor order parity, existing row-shuffle parity, nested/null data, bounded input consumption and reader counts, Arrow offset overflow, empty input, filters/limits, vector tensors, deletions, historical snapshots and persistent spawn workers across two explicit ranks.
• Repository-configured Flake8, license checks, Python 3.6 grammar checks and git diff --check passed.