Add reproducible seeded submission order

[xujustinj]

A sorted dataset can cluster similar documents during concurrent inference. Add optional --shuffle_seed to the benchmark and inference commands so comparisons share a reproducible order independent of filesystem discovery. Save the seed and ordered IDs in _submission_order.json.

Schedule async tasks in input order before observing completion order. Omitting the seed preserves discovery order; cached results are still skipped normally.

Validation:

• 9 focused tests: shuffle reproducibility across pipelines/discovery order, different seeds, unchanged unseeded order, async scheduling, CLI forwarding, existing external filename behavior.
• Repository-wide Ruff lint and format checks.
• git diff --check.
• No provider requests or paid inference.