Round-trips the multimodal pipeline from the embedded chat:
* 📎 button beside the textarea opens a file picker (`accept="image/*"`),
* the selected image is read as a data URL, the base64 payload is
  stripped off the `data:image/...;base64,` prefix for the wire format
  and a thumbnail preview bar appears above the textarea until the
  user sends or removes it with ×,
* image-only turns are allowed (the model falls back to its default
  "describe this image" behaviour); fully empty submits are not,
* the user bubble renders the same thumbnail above the prompt so the
  conversation transcript stays self-contained.

Dispatch logic — `send()` picks the endpoint based on `pendingImage`:
* with image  → POST /api/chat (Ollama NDJSON), body shape
  `{ messages: [{ role:"user", content, images:[<base64>] }], stream:true }`.
  History is intentionally NOT replayed: the multimodal MVP is a one-shot
  probe (matching the backend), and re-sending old base64 bytes every
  turn would balloon the prompt context for nothing.
* without image → POST /v1/chat/completions (SSE), unchanged.
* the streaming loop handles both formats from the same `while/read`:
  NDJSON lines parse straight as JSON with `message.content`; SSE lines
  keep their `data:` prefix detection and `choices[0].delta.content`.

Attachment is cleared on send and on "clear conversation" so a stale
image can never leak into a later turn.

Releasing the web-chat multimodal MVP:
  * /api/chat routes the Ollama `images:[<base64>]` field through
    engine.generate_multimodal() when an mmproj is loaded,
  * loading a multimodal model now resolves the mmproj sibling from
    the store and forces sequential mode (the batching scheduler is
    text-only),
  * the embedded chat ships an attach-image button + thumbnail preview
    and dispatches multimodal turns over /api/chat (NDJSON) while
    text-only turns keep using /v1/chat/completions (SSE).