Backport multimodal tool outputs and LiteLLM content sanitation

[bitifirefly]

Summary

• backport multimodal tool-output flow so read_file and web_fetch can return native image content blocks
• add shared image helpers (detect_image_mime, build_image_content_blocks) and use them across tools
• enrich system/subagent prompts to explicitly allow visual-resource reading via tool outputs
• keep tool interfaces compatible with structured results (Tool.execute / ToolRegistry.execute -> Any)
• strip internal _meta keys from multimodal content blocks before LiteLLM requests to avoid provider schema incompatibilities

Testing

• ruff check opencane/utils/helpers.py opencane/agent/tools/filesystem.py opencane/agent/tools/web.py opencane/agent/context.py opencane/agent/subagent.py opencane/agent/tools/base.py opencane/agent/tools/registry.py opencane/providers/litellm_provider.py tests/test_filesystem_tools_workspace.py tests/test_web_fetch_security.py tests/test_litellm_message_sanitize.py
• pytest -q tests/test_filesystem_tools_workspace.py tests/test_web_fetch_security.py tests/test_litellm_message_sanitize.py
• pytest -q (431 passed)