fix(asyncio): keep pipeline packing off the event loop

[lh0156]

Summary

Large async pipelines currently call Connection.pack_commands() synchronously from the event loop. Packing a large transaction or pipeline can therefore block unrelated coroutines for seconds.

This change:

• offloads command packing for transactional and non-transactional async pipelines with asyncio.to_thread
• preserves the existing wire packing and response handling
• adds a deterministic regression test covering both pipeline paths

Fixes #3416.

Validation

• TDD regression test was red before the production change and green after it
• pytest -q tests/test_asyncio/test_pipeline.py -k 'pipeline_packing_does_not_block_event_loop or TestAsyncPipelineOperationDurationMetricsRecording' — 7 passed
• RESP3 async pipeline suite — 71 passed, 2 skipped
• RESP2 async pipeline suite — 71 passed, 2 skipped
• invoke linters — passed

The full Docker/fixed-client matrix was not completed because the local Docker credential helper and required cluster/maintenance services were unavailable.

[petyaslavova]

Thanks for tackling this — the underlying issue in #3416 is real: for large async pipelines/transactions, pack_commands runs synchronously on the event loop and can block it for seconds, and moving packing off the loop is a reasonable direction.

There is, though, one design point to settle: the offload is applied unconditionally, so every async pipeline — including small ones — now pays a to_thread dispatch and uses the loop's shared default executor.
It will be better to gate the offload behind a batch-size threshold (so the common small-pipeline path keeps its current cost) or share a quick benchmark showing the small-pipeline overhead is negligible.

It would also help to note whether async cluster pipelines (redis/asyncio/cluster.py, which pack the same way) are in scope here or a follow-up.

Once that's decided and the PR is out of draft, it should be ready for another review.