feat(sn-da-image-caption): parallelize caption_batch via --concurrency

[gangbangxie]

Problem

caption_batch() in skills/sn-da-image-caption/scripts/caption.py loops through a directory one image at a time:

for i, img_path in enumerate(image_files, 1):
    result = caption_image(img_path, ...)

Meanwhile the sister script skills/sn-ppt-entry/scripts/caption_images.py already parallelizes its VLM calls with ThreadPoolExecutor. For a directory of 50 images at ~2-4s each, the standalone path costs 100-200s end-to-end — which is what most people hit when running large batches.

Issue: #92.

Change

• New --concurrency / -j flag, default 1, clamped to 1-16 (matches the bounds used in caption_images.py).
• When concurrency > 1, calls go through a ThreadPoolExecutor and results are written back into a pre-sized list by index, so the returned order matches the sorted file order regardless of completion order.
• When concurrency == 1 (default), behavior is byte-for-byte the same as before: same loop, same caption_image calls. The serial output line still prints index + filename + status.
• caption_image() is already thread-safe — each invocation builds its own OpenAI client, the cache writes are keyed by <md5>_<prompt_hash>.txt so different threads write to different files, and os.makedirs(..., exist_ok=True) is race-safe.

Diff is +33/-8 in one file.

Testing

Local smoke test against a mocked caption_image that sleeps 50-200ms per call:

• 8 simulated images, -j 1: ~0.80s, original output style preserved
• 8 simulated images, -j 4: ~0.33s (≈2.4× speedup)
• Returned list has identical order in both modes
• -j 0 and -j 999 clamp safely to 1 and 16
• Empty directory returns []
• python3 caption.py --help still parses cleanly

Happy to adjust the default (e.g. raise to 4 to match caption_images.py) if you'd prefer — left at 1 so this is a pure opt-in change.

Closes #92