feat(vision): image + screen input for jarvis ask

[joshazmy]

What does this PR do?

Adds image input to jarvis ask so OpenJarvis can use vision-capable local
models (e.g. gemma3:4b) without giving up its local-first promise.

• -i / --image <file> — attach one or more images (repeatable).
• -S / --screen — capture the primary monitor and send it.
• Message.images (base64) → messages_to_dicts() → Ollama /api/chat
images. Text-only requests are unchanged.
• Vision auto-routes to direct-to-engine mode; with an explicit --agent it
  warns instead of silently dropping the image.
• Privacy guard warns before sending an image to a non-local engine; the
  security guardrail preserves images when sanitizing a flagged prompt.
• Adds JARVIS_NUM_CTX (default 16384) to size the Ollama context window
  for an image plus a conversation.

Closes #485.

How was this tested?

• 6 new unit tests in tests/test_vision.py covering the data-flow contract:
  image forwarding to the serializer, text-only messages unaffected, the
  guardrail preserving images through sanitization, and JARVIS_NUM_CTX
  parsing (default + override + invalid-fallback). All pass locally.
• Manually verified end to end on Windows with Ollama + gemma3:4b (both
  --image and --screen), running 100% on-GPU.
• Honest caveat: the non-Windows screen-capture path (mss / Pillow) is
  implemented but I have only verified the Windows .NET path on real
  hardware — a macOS/Linux check would be appreciated.

Checklist

• Tests pass — new tests/test_vision.py passes locally; the full suite
  runs on CI (my local venv lacks some optional engine test deps, e.g.
  respx).
• Linter passes (uv run ruff check src/ tests/)
• Formatter passes on the changed files (ruff format --check). I left the
  78 unrelated pre-existing files alone — they differ only because of a ruff
  version mismatch, not this change.
• New/changed public API has docstrings
• Follows registry pattern — N/A; this wraps the existing engine/CLI and
  adds no new registry component.
• Documentation updated — docs/user-guide/cli.md (new "Vision Input"
  section + flag table) and CHANGELOG.md.