[DOCS] Read attaches image files for vision models #1024

Summary

The CLI's Read tool now reads an image file as an attachment for a model that can see images (inference-gateway/cli#1599). Before, Read turned every image file away for every model, and the only way to look at a file on disk was ImageDecode, even for a vision model.

What changed, user-facing:

• Read on a png, jpg, jpeg, gif or webp file, or on a file without an extension whose bytes are an image, attaches the image to the model's next message when the gateway lists image input among the model's modalities. The tool result carries the file's path and pixel size.
• A text-only model, or a session without an image service, gets an error from Read that points at ImageDecode, which describes the image in words through the vision annotator.
• ImageDecode itself is unchanged and still works for every model.

The docs say the opposite today, in two places:

• cli.md, the "Vision-capable models and ImageDecode" section (around line 1964) and the frame sources section (around line 1568, "ImageDecode handles arbitrary images"): a vision model now reads a file on disk with Read, ImageDecode is for a text-only model or for a text description on top of the image.
• github-action.md, around line 495: the guidance that an agent downloads an embedded image with WebFetch and then reads it with ImageDecode. With a vision model it reads the downloaded file with Read, and ImageDecode stays the path for a text-only session model with vision-model set.

Source: inference-gateway/cli#1599.

Acceptance Criteria

• cli.md explains that Read attaches an image file for a vision model and sends a text-only model to ImageDecode, with a one-line example of each outcome.
• cli.md's frame sources and "Vision-capable models and ImageDecode" sections no longer say ImageDecode is the way to read an arbitrary image for every model.
• github-action.md's image guidance names Read for a vision session model and ImageDecode for a text-only one with vision-model set.