fix(slack): restore gap image DataContent in thread hydration

[Aaronontheweb]

Summary

• The prior "drop gap image bytes" optimization (merged in feat(slack): backfill thread history on mid-thread mention #576) stripped images from hydrated thread history unconditionally. This broke vision on prior-thread context for image-capable models — exactly the feature the spec documents with "Backfilled messages SHALL include inline multimodal content (images as DataContent) so the LLM receives them as vision content."
• The modality decision belongs to the session actor, not the channel binding actor. `LlmSessionActor` already has a capability gate (`src/Netclaw.Actors/Sessions/LlmSessionActor.cs:1620`) that strips image media refs and notifies the user when the resolved model can't handle vision.
• This restores the original merge behavior: gap `DataContent` items flow through to the session, the session's gate decides based on `IModelCapabilityResolver`.

Test plan

• `dotnet build`
• `dotnet slopwatch analyze` — 0 issues
• `dotnet test` — all 122 Slack tests pass, including the restored `Backfill_messages_are_merged_into_single_user_turn` assertion that gap `DataContent` reaches the session.