Multimedia Support: Nexa SDK integration

[iwr-redmond]

The Nexa SDK includes easy to use classes for locally based text and image generation.

Adopting this SDK for locally based generations would allow for:

• More reliable setup, versus the unmanaged Ollama dependency
• Image generation by LLM agents using models such as Flux Schnell
• Image 'viewing' by LLM agents using models such as OmniVLM (e.g. the output from 'describe this image in detail' could be passed to the main agent LLM)
• Validated models to be easily distributed at runtime to users using the nexa.general.pull_model function (e.g. specified as part of the scenario YAML)

[iwr-redmond]

Addendum: once nexa-sdk#370 is complete, you would also be able to reliably generate JSON responses from small LLMs without errors using token-level supervision. There is already a prototype of this feature (nexa-sdk#368) by @xsxszab but it is not especially robust.