Prevent image attachments on models that don't support image input

[bilthon]

Summary

Today the chat composer lets you attach an image to any model, including text-only ones. Sending that request hits the upstream provider, fails (the model can't accept image input), and still costs sats. This PR gates image attachments to vision-capable models across every entry path, while keeping PDF uploads working everywhere (PDFs are extracted to text client-side and are valid input for text-only models).

A model is treated as image-capable when any of its architecture.input_modalities normalizes to "image" (via the existing normalizeModality helper.

Changes

• The image/* type hint is removed from the accept attribute of the input tag when a model incapable of processing images is selected, thus preventing the user from selecting an image.
• Post processing logic removes the file and displays an alert in case the user forces an image or drag-and-drops one.
• Image files are removed in case the user switches from an image-capable model to one that can't handle them, in case there were image files already selected.

Closes #171