…text per request

Three independent correctness fixes for the web multimodal path, all
verified against today's QA logs:

1) WebP / AVIF / HEIC images (ui/app.js)
   mtmd decodes via stb_image (jpg/png/bmp/gif only). WebP reached the
   server as `mtmd_helper_bitmap_init_from_buf: failed to decode image
   bytes`, the stream ended at 0 chunks, and the model then hallucinated.
   The picker now re-encodes any non-stb format to PNG in-browser via a
   canvas, transparently. HEIC (which the browser itself can't decode)
   gets an actionable message instead of silence.

2) Surface backend stream errors (ui/app.js)
   The server already emits `{"error": "..."}` as an NDJSON line, but the
   UI only looked for `message.content`, so decode failures showed as an
   empty "0 chunks" reply that looked like a hallucination. The stream
   loop now renders the error in the assistant bubble and stops.

3) EXPERIMENT — rebuild MtmdContext per request (inference/mod.rs)
   The image is encoded every turn (logs show 273/260 tokens) yet the
   model is intermittently "blind" and confabulates (repeated-text in
   Chinese/Cyrillic). The only mutable state shared across the otherwise
   fresh per-request LlamaContext is the long-lived MtmdContext, so we
   now rebuild it per request to rule that out. If the flakiness
   disappears the bug was shared mtmd state; if not, it points upstream
   (mtmd_helper_eval_chunks / SWA KV-position handling, llama.cpp#17930)
   and we stop chasing it from our side.