feat: add end-to-end multimodal image turn flow with view_image fallback and persistence redaction

[wgqqqqq]

Summary

This PR upgrades the image handling pipeline across the agentic chat flow.

It replaces the old image-analysis tool path with view_image, enables direct
image attachment for multimodal-capable models, keeps a vision pre-analysis
fallback for text-only models, and redacts inline image payloads from
persisted session data.

What Changed

• replaced the legacy image analysis tool with view_image
• extracted shared image-processing and vision-model resolution utilities
• added multimodal user message support in the core message model
• enabled direct image turn input when the selected model supports image
  understanding
• kept pre-analysis as a fallback path for text-only primary models
• added temporary uploaded image context resolution for clipboard/image cache
  flows
• sanitized persisted messages and context snapshots to remove stored
  data_url payloads
• extended AI connection testing to verify multimodal image-input capability

Details

Tooling and image pipeline

• migrated from the old analyze_image tool implementation to view_image
• centralized image loading, path resolution, mime detection, optimization,
  and vision-model lookup
• updated tool registration and related frontend tool-card wiring

Multimodal turn support

• added imageContexts support to the dialog-turn API
• allowed the frontend sender to choose between:
  • direct attach for multimodal-capable models
  • vision pre-analysis for text-only models
• uploaded clipboard images into a temporary cache so turn requests can
  resolve missing image payloads safely

Execution and model integration

• added multimodal user message construction in core messaging
• updated execution flow and tool pipeline state to carry image context
  through the round lifecycle
• added image-aware token estimation and client-side multimodal request
  support

Persistence and safety

• redacted base64/data URL image payloads before writing messages or turn
  context snapshots to disk
• preserved useful image metadata for replay/debugging without storing the raw
  image content

cc @wsp1911