fix: default to get_num_physical_cores for model loading

For consistency, since std::thread::hardware_concurrency can
return double the number of cores on SMT machines.

Author: wbruna

model.cpp

@@ -1956,7 +1956,7 @@ bool ModelLoader::load_tensors(on_new_tensor_cb_t on_new_tensor_cb, int n_thread
     std::atomic<int64_t> copy_to_backend_time_ms(0);
     std::atomic<int64_t> convert_time_ms(0);
 
-    int num_threads_to_use = n_threads_p > 0 ? n_threads_p : (int)std::thread::hardware_concurrency();
+    int num_threads_to_use = n_threads_p > 0 ? n_threads_p : get_num_physical_cores();
     LOG_DEBUG("using %d threads for model loading", num_threads_to_use);
 
     int64_t start_time = ggml_time_ms();