Add vision caption console and encoder pipeline

[zhongkaifu]

Summary

• add a VisionCaption application that turns image patches into transformer-friendly embeddings so we can train captioning models without the legacy Image2Seq code
• expose the new application through a dedicated VisionCaptionConsole CLI and wire it into the solution
• fix the vision corpus token accounting and add ImageSharp as a dependency for loading and normalizing images

Testing

• ⚠️ dotnet build Seq2SeqSharp.sln (not run: dotnet CLI is unavailable in the execution environment)

---

Codex Task