Add moondream3 model for CUA

[ddupont808]

This PR adds support for the moondream3 model in the Agent SDK. The moondream3 model supports both UI screenshot parsing (element detection + captioning) and UI grounding (element localization). As a result, it also enables support for composed agents using text-only models.

The model can be used in a few different ways:

Grounding only

agent = ComputerAgent("moondream3")
agent.predict_click(screenshot, "start button") # ( x, y )

Composed Agent (Supports vision & text-only models)

agent = ComputerAgent("moondream3+ollama/gemma3")
agent.run("close this window")

This will caption each screenshot sent to the model and append a user message with a list of elements.

Example output:

Detected form UI elements on screen:

• Full Name
• Email Address
• Priority Level
• Subject
• Detailed Description
• Request Support