feat(closes OPEN-10341): add native async runner for testset batches

[gustavocidornelas]

Pull Request

Summary

Adds native async support to OpenlayerModel.run_batch_from_df so that customers whose per-row work hits slow APIs
(~5s/row) can run testset batches concurrently instead of strictly sequentially.

Users opt in by defining run as async def run(...); the framework then drives rows through asyncio.gather gated by a semaphore. Sync run keeps today's behavior byte-for-byte.

Changes

• run may now be defined as async def run(...). When it is, run_batch_from_df dispatches rows concurrently
  via asyncio.gather + asyncio.Semaphore(max_workers).
• New max_workers kwarg on run_batch_from_df and batch, plus --max-workers on the CLI. Default resolves to
  4 for async run, 1 for sync run — writing async def is the opt-in signal that interleaving is safe.
• max_workers > 1 with a sync run raises ValueError rather than silently ignoring it.
• Per-row exceptions still propagate and abort the batch (fail-fast, same as today). For the async path,
  asyncio.gather cancels in-flight siblings before re-raising.
• Extracted _run_rows_async, _apply_row_result, and _build_config helpers so the row bookkeeping is shared
  between the sync and async paths.

Context

OPEN-10341: Add native async runner for testset batches

Testing

• Manual testing

Monitoring

• No expected impact

Notes

• Backwards compatible: existing sync run implementations behave identically. Same sequential code path, same
  fail-fast semantics, no executor or asyncio overhead.
• Customer-facing usage: if a customer's openlayer_run.py already defines async def run(...), they get 4-way
  concurrency automatically next release. To override, append --max-workers N to the batchCommand in
  openlayer.json.