Add MLX multimodal inference support

[jbg]

Summary

• update the safemlx dependency family to the latest compatible releases
• enable image processing for MLX models and route ordered image/text inputs through model processors
• preserve text-only generation and disable draft-model generation when media is present
• require processor metadata when advertising supported multimodal MLX repositories
• add focused coverage for prompt ordering, processor requirements, marker validation, and text-only behavior

Why

Recent safemlx releases add multimodal model processors and change prepared-input generation APIs. Goose needs to use those APIs so MLX vision models receive decoded image data alongside the rendered prompt while text-only models continue using the existing token input path.

Validation

• cargo fmt
• cargo check -p goose-local-inference --features mlx
• cargo test -p goose-local-inference --features mlx (130 tests passed)
• cargo check -p goose-local-inference

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 29a59a24a8

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[chatgpt-codex-connector]

Account for media tokens before capping generation

When has_media is true, this branch prepares a processor input that can expand each image into additional model positions, but the earlier context check and the later mlx_max_tokens(..., prompt_tokens.len()) still use the plain text token count from the marker-only prompt. For image requests near the model context limit, Goose can allow a prompt that only fits before image expansion and then either fail during generation or overrun the intended context budget; use the prepared input's effective length or an image-token estimate when validating and computing headroom.

Useful? React with 👍 / 👎.