Use MmapUseMlockIgnoreErrors to pin model pages in RAM

On memory-constrained machines (e.g. 32 GB with a 21 GB model), plain
Mmap causes constant page eviction and re-fault — 0.2 tok/s decode.
MmapUseMlockIgnoreErrors calls mlock() to pin pages after mmap, keeping
weights resident once loaded. Falls back silently if mlock exceeds the
system limit.  Measured 4.7 tok/s decode on M3 Pro 32 GB (24x speedup).

Co-authored-by: Claude <noreply@anthropic.com>

Author: mergennachin

examples/models/gemma4_31b/main.cpp

@@ -155,7 +155,7 @@ int main(int argc, char** argv) {
   auto module = std::make_unique<Module>(
       FLAGS_model_path,
       data_files,
-      Module::LoadMode::Mmap,
+      Module::LoadMode::MmapUseMlockIgnoreErrors,
       /*event_tracer=*/nullptr,
       /*memory_allocator=*/nullptr,
       /*temp_allocator=*/nullptr,