The mmap region kept in system RAM for -ncmoe can be tens of GB, larger than
device->max_buffer_size (~4 GiB), so importing it as a single Vulkan buffer
failed and pinning silently no-op'd. Import in page-aligned chunks up to
max_buffer_size, registering each in device->pinned_memory. Add a bound check
in the pinned copy path so a tensor that straddles a chunk boundary falls
back to staging instead of reading out of bounds.
Assisted-by: opencode
Lumpiasty
99f8050c32