feat(engine): env-var switch for async-first models experiment

[eric-tramel]

Summary

Opt-in async execution path for the cell-by-cell generation pipeline, gated behind DATA_DESIGNER_ASYNC_ENGINE=1. Replaces thread-pool concurrency with native asyncio for LLM calls while keeping the sync path as the unchanged default.

What changed

• Async ModelFacade methods — acompletion(), agenerate(), agenerate_text_embeddings() on ModelFacade, using LiteLLM Router's native async APIs. MCP tool calls offloaded via asyncio.to_thread().
• acatch_llm_exceptions — async mirror of the existing sync decorator, sharing the same error-mapping logic.
• AsyncConcurrentExecutor — drop-in async replacement for ConcurrentThreadExecutor. Uses TaskGroup + Semaphore for bounded concurrency, asyncio.Event for early shutdown. Runs on a persistent background event loop (same pattern as mcp/io.py) to avoid breaking libraries that maintain internal async state across calls.
• ColumnWiseBuilder — branches on env var: _fan_out_with_async() vs existing _fan_out_with_threads().
• agenerate() on generators — async path in ColumnGeneratorWithModelChatCompletion.
• Benchmark — updated with async mock support for --mode compare.

Benchmark results

With mock LLMs (zero-latency stubs), sync and async produce identical output (SHA256 content hash match across all iterations). No measurable speedup is expected here — the async path's advantage comes from real network-bound LLM calls where await yields the event loop during I/O wait. With mocks there's no I/O to overlap.

Validated end-to-end with the pdf_qa recipe using real inference (Opus 4.5 + MCP tool calls).

How to test

# Benchmark (correctness)
uv run python scripts/benchmarks/benchmark_engine_v2.py --mode compare --iterations 3 --num-records 64

# Existing tests (no regressions)
uv run pytest packages/data-designer-engine/tests -x -q

# Real inference
DATA_DESIGNER_ASYNC_ENGINE=1 uv run python docs/assets/recipes/mcp_and_tooluse/pdf_qa.py --model-alias <your-model> --num-records 2

Test plan

• 1086 engine tests pass
• Benchmark content match: identical hash sync vs async
• pdf_qa recipe completes with real inference, no event loop errors
• Async path confirmed via log markers (one per LLM column)