On a device without mmap support, --lazy-mode auto resolved to off even
with --load-mode mmap, so llama_prefetch_rows never ran and the rows of
the large tensors faulted in one page at a time. Resolve it to off only
when mmap is off; the default iGPU load path is unchanged.

Fixes #31943