Proposal: permission-aware image analysis tools example

[somnus-J-307]

Problem

Users who want an agent to perform deterministic operations on a local image
need a reference for connecting image-processing functions to AgentScope tools.
The example should show how to keep file access scoped, request permission
before saving derived images, and return measurements that the agent can report
without inventing visual conclusions.

Proposed example

Add a focused examples/image_analysis/ example that:

1. Registers a read-only grayscale-inspection tool and one output-producing
   operation through FunctionTool and Toolkit.
2. Accepts only images in a configured input directory and saves derived PNGs
   in a configured output directory, using AgentScope's permission flow for
   writes.
3. Provides a small synthetic image and a no-API-key demo that verifies tool
   calls, measurable results, and an error for a missing image.
4. Documents how to run the same tools with a real model, what the output means,
   and what has not been validated.

The current local prototype supports 8-bit single-frame images, ROI selection,
gray-level statistics, Otsu segmentation, and Sobel edge maps. For the main
repository, I would reduce this to the smallest set of operations that best
demonstrates AgentScope's tool and permission behavior. Pillow would be an
example-only dependency; NumPy is already a core AgentScope dependency.

Alternatives considered

• Put the full image-processing application in agentscope-samples and link to
  it from documentation. This seems more suitable if maintainers prefer the
  main repository's examples to remain very small.
• Add image-processing functions to the core package. I do not propose this:
  the algorithms are application-specific and do not need a new framework API.

Validation available

The local prototype has six unit tests covering known image measurements,
ROI boundaries, unsupported depth, file scope, and transparent pixels. An
offline scripted-model run exercises the actual AgentScope agent/tool loop.
The interactive real-model path has only been checked for startup and missing
credentials; its conversation behavior has not been verified with a provider.

Would an examples/image_analysis/ PR be welcome in the main repository, or
would agentscope-samples be the better home? Which single output-producing
operation would be most useful to retain in a concise example?

Originally proposed in #2814; moved here following the automated triage suggestion.