WIP: Add support for CogAgent

[Tianyue-Zhao]

Overview

This PR adds support for CogAgent
CogAgent is a visual model specializing in GUI recognition and visual grounding (providing accurate pixel coordinates for GUI actions)
It is based on the earlier CogVLM vision model, but has improved high-res performance and visual grounding for use as a GUI agent

Completes #4387

Architecture

• CogAgent uses two CLIP encoders
• The first one is largely similar to existing vision encoders in llama.cpp (LLAVA)
  • Therefore the first encoder fits right in to the new infrastructure in llama : second attempt to refactor vision API #11292
• The second encoders is for high-res images (1120 x 1120)
  • Instead of using visual tokens, it transfers info through a cross-attention mechanism to the main LLM
  • I'm not yet sure where to put this in the final implementation, since the vision refactor doesn't really provide facilities for double-encoder vision models

Current status

The current implementation is locally tested but temporary and not ready for usage
This PR waits for the new vision infrastructure brought by #11292
When the vision infrastructure in that PR is finalized, I will update the CogAgent implementation to take advantage of the new features