fix: reject invalid n_dims in tensor header to prevent stack-buffer-overflow

[shafiuzzaman-md]

Problem

bark_model_load() (bark.cpp:1019) reads a tensor header's n_dims field as a raw int32_t directly from the model file and uses it, unbounded, as the loop count writing into a fixed-size 2-element stack array int32_t ne[2]:

int32_t nelements = 1;
int32_t ne[2] = {1, 1};
for (int i = 0; i < n_dims; ++i) {
    read_safe(fin, ne[i]);   // n_dims is fully attacker-controlled, no upper bound
    nelements *= ne[i];
}

A model file with n_dims > 2 writes past the ne[2] array on the stack. Reproduced with a crafted 144-byte model file (n_dims = 20) through the stock main -m <file> -p "text" command line, under an AddressSanitizer build:

==...==ERROR: AddressSanitizer: stack-buffer-overflow
WRITE of size 4 at ... thread T0
    #0 memcpy
    #1-3 std::istream::read / filebuf / streambuf
    #4 read_safe<int>          bark.cpp:168
    #5 bark_model_load         bark.cpp:1019
    #6 bark_load_model_from_file  bark.cpp:1124
    #7 bark_load_model         bark.cpp:1174
    #8 main                    examples/main/main.cpp:55

The identical copy-pasted loop also exists in ggml_quantize_weights (the offline quantize tool, bark.cpp:357, int32_t ne[4]), fixed here too.

Fix

Add an explicit n_dims range check immediately after it is read, before the loop that populates ne[], at both sites sharing this pattern:

if (n_dims < 1 || n_dims > 2) {   // 4 for ggml_quantize_weights
    fprintf(stderr, "%s: invalid n_dims %d in model file (expected 1 <= n_dims <= 2)\n", __func__, n_dims);
    return false;
}

For normal, well-formed model files this is a no-op; a malformed file is now rejected with a clear error message instead of corrupting the stack.

Testing

• Before: the crafted model file (n_dims=20) crashes with an ASan stack-buffer-overflow WRITE at bark_model_load (bark.cpp:1019).
• After: the same file is rejected cleanly ("invalid n_dims 20 in model file"), no crash.
• Regression check: a file with a valid n_dims (2) passes the new check unaffected (fails later only because the synthetic test file has no real tensor data, unrelated to this fix).
• Both main and quantize targets build cleanly with the change.

[shafiuzzaman-md]

The 3 failing checks (ios, emscripten, ubuntu-latest-run) are unrelated to this change. Each fails before reaching the modified code:

• ios: fatal error: 'ggml-alloc.h' file not found in the Xcode SwiftUI example project's include paths.
• emscripten: error: "embind requires -std=c++17 or newer", the emscripten build isn't passing -std=c++17.
• ubuntu-latest-run: fails at the apt-get dependencies step (Unable to locate package python3.10-venv), before build/run even starts.

The two checks that actually compile and run the changed code (windows-msys2 CLANG64 and UCRT64) both pass. Happy to help fix the CI separately if useful, but wanted to flag that these failures predate and are unrelated to this PR.