Docs: OpenAI vision how-to miscounts API requests and its snippet drops version= #73

The OpenAI vision how-to miscounts the example's API requests, and its snippet
omits the explicit version= that the same page recommends.

Current behavior

The request count is wrong

docs/how-to/call-openai-vision.md:94-98 says the example makes one model
request per episode. The example ships two checks, and each check makes its own
client.responses.create call:

• describe_activity at examples/openai_vision/pipeline.py:42-76
• hand_visibility at examples/openai_vision/pipeline.py:79-119

An episode therefore makes one request per check, or two requests for the
shipped example. The fps=0.5 and max_tiles=12 values in the same paragraph
are correct.

The snippet omits version=

The snippet at docs/how-to/call-openai-vision.md:67 uses
@app.check(uses="vision"). The corresponding example at
examples/openai_vision/pipeline.py:42 passes
version="responses-contact-sheet-v1". The prose at
docs/how-to/call-openai-vision.md:96-98 also tells you to version a step when
external model configuration cannot be hashed from the function.

Add an explicit version= to the snippet. Keeping uses="vision" instead of
introducing the example's alias constant is fine.

Pattern to copy

Copy the decorator shape from examples/openai_vision/pipeline.py:42 and keep
the concise explanatory style of the surrounding how-to.

Definition of done

• The request-count paragraph says one request per check and two per episode
  for the shipped example.
• The snippet passes an explicit version=.
• The existing frame sampling and tile-cap values remain unchanged.

Validation

lychee --no-progress --include-fragments \
  --exclude '^https://github\.com/Hebbian-Robotics/hflow/(issues|security/advisories/new)$' \
  --exclude-path references/mcap-spec.md \
  --exclude-path references/foxglove-CompressedVideo.proto .