…ention)

The 1024-token crash exposed the real cap. Gemma's vision encoder uses
non-causal attention, which requires the whole image in one micro-batch:
GGML_ASSERT(causal_attn || n_ubatch >= n_tokens). Our n_ubatch defaulted to
512, so:
  * images > 512 tokens hard-abort (the EULLM_IMAGE_MAX_TOKENS=1024 SIGABRT),
  * and effective image resolution was silently capped at 512 tokens
    regardless of image_max_tokens — which is why 280 vs 512 gave identical
    output: the model never actually received more than ~512 image tokens.

Fix: in the multimodal context only, set both n_batch and n_ubatch to
max(config.n_batch, EULLM_IMAGE_MAX_TOKENS, 512). Now raising the image
budget genuinely increases resolution instead of crashing, and we can finally
test whether more resolution fixes the hard (dark / portrait) images. Text
and scheduler contexts are untouched (no VRAM regression there).