Fix for hanging on model load on Windows

[DmitryMalishev]

The is the fix for the issue, when integration test for "qvac-lib-infer-llamacpp-llm" on Windows hangs.

This small missprint "ro" instead of "rb" caused undefined behavior on Windows: on the second opening a model file it always got identical file ID as was used for the first open. But fclose() already called at first open, so that second descriptor had not-null but still invalid ID, so that it hanged on fseek() method to detect file size.

Sorry, string numbers are incorrect bellow, as I added many logs and they are counted in my local repo state.

Calling from Llm-Addon
  └─> llama_model_load_from_file()                    [llama.cpp/src/llama.cpp:256]
       └─> create_disk_fileloader()                    [llama.cpp/src/llama.cpp:250]
            └─> llama_model_loader::llama_model_loader() [llama.cpp/src/llama-model-loader.cpp:503]
                 └─> gguf_file_load::gguf_file_load()   [llama.cpp/src/llama-model-load.cpp:10]
                      │
                      ├─> [FIRST OPEN PATH]
                      │   gguf_init_from_file()          [llama.cpp/ggml/src/gguf.cpp:806]
                      │    └─> ggml_fopen(fname, "rb")   [llama.cpp/ggml/src/ggml.c:560]
                      │         └─> _wfopen()             [Windows CRT]
                      │              ├─> Returns: FILE* = 00000231E523F1D0
                      │              └─> fd = 3
                      │    
                      │    gguf_init_from_file_impl()    [llama.cpp/ggml/src/gguf.cpp:799]
                      │     └─> Reads metadata from FILE*
                      │    
                      │    fclose(FILE* 00000231E523F1D0) [llama.cpp/ggml/src/gguf.cpp:820]
                      │     └─> Returns FILE* to CRT pool
                      │    
                      │    Sleep(50)                      [llama.cpp/ggml/src/gguf.cpp:826]
                      │     └─> Waits 50ms
                      │
                      └─> [SECOND OPEN PATH - HANGS HERE]
                          llama_file_disk constructor     [llama.cpp/src/llama-model-load.cpp:31]
                           └─> std::make_unique<llama_file_disk>(fname, "rb")
                                └─> llama_file_disk::llama_file_disk() [llama.cpp/src/llama-mmap.cpp:250]
                                     └─> pimpl(std::make_unique<impl>(fname, mode))
                                          └─> impl::impl()              [llama.cpp/src/llama-mmap.cpp:180]
                                               └─> ggml_fopen(fname, "rb") [ggml.c:560]
                                                    └─> _wfopen()          [Windows CRT]
                                                         ├─> Returns: FILE* = 00000231E523F1D0 ⚠️ REUSED!
                                                         └─> fd = 3 ⚠️ SAME FD!
                                               
                                               ✅ fp = 00000231E523F1D0
                                               ✅ Check fp != NULL passes
                                               
                                               seek(0, SEEK_END)         [llama-mmap.cpp:191]
                                                └─> _fseeki64(fp, 0, SEEK_END)
                                                     └─> ⚠️ **HANGS HERE** ⚠️
                                                         Operating on reused/corrupted FILE*

[jesusmb1995]

Thanks for finding this bug and fixing it.

Indeed, if you look at upstream Llama.cpp it uses "rb":

    files.emplace_back(new llama_file(fname.c_str(), "rb"));

[jesusmb1995]

This PR also fixes OpenBLAS CI 🙂 , See successful CI https://github.com/tetherto/qvac-ext-lib-llama.cpp/actions/runs/18656998019/job/53188454321?pr=40

Please help review/merge @jpgaribotti @olyasir