fit: clear use_direct_io on dry-run model load for params fitting

O_DIRECT on a 100GB model file during the metadata-only scan hangs.
The dry-run only needs metadata so normal buffered IO is fine.

Author: kmbandy

common/fit.cpp

@@ -53,9 +53,10 @@ static std::vector<llama_device_memory_data> common_get_device_memory_data(
     }, &ud);
 
     llama_model_params mparams_copy = *mparams;
-    mparams_copy.no_alloc  = true;
-    mparams_copy.use_mmap  = false;
-    mparams_copy.use_mlock = false;
+    mparams_copy.no_alloc       = true;
+    mparams_copy.use_mmap       = false;
+    mparams_copy.use_mlock      = false;
+    mparams_copy.use_direct_io  = false;
 
     llama_model * model = llama_model_load_from_file(path_model, mparams_copy);
     if (model == nullptr) {