Compared our generate_multimodal against llama.cpp's reference mtmd-cli
(tools/mtmd/mtmd-cli.cpp). The reference sets `text.add_special = add_bos`
(true on the first turn), so it prepends the model's BOS token. We set
`add_special = !request.raw`, and the multimodal path always uses raw=true
(hand-built turn template) — so we were sending the prompt WITHOUT <bos>.

Gemma requires BOS. Without it the prompt is malformed and the model
degrades into "wall of text / line art / collage" confabulation — but only
when the image signal is weak. Strong landscapes survived the missing BOS
(which is why some images "worked"); portraits, people and dense/abstract
images tipped into garbage. This reproduced identically on CLI and web, and
on 0.6.0 — i.e. it was never a beta regression nor the web pipeline; it was
this one missing token in the shared core, present since multimodal landed.

Fix: add_special = true (always prepend BOS) for the mtmd tokenize, matching
the reference. parse_special stays true so the hand-built turn markers are
still recognised.

The headline of this beta: the missing-<bos> fix in the image prompt, which
reproduced the "wall of text / line art" misreads on CLI, web and 0.6.0
alike. Also carries the n_ubatch sizing, image_max_tokens override, WebP→PNG,
error surfacing, web audio, and the local CUDA installer CI change.

First release on `method: local` for the Windows CUDA toolkit — that step is
slow once (downloads + caches the full offline installer), fast afterwards.