[CPU EP] MatMulNBits (accuracy_level=4) produces wrong results on x86-64 AVX2 when loaded from session.save_external_prepacked_constant_initializers output (1.31.1) #33677

Describe the issue

We save an optimized model with prepacked constant initializers and load that saved model later
(the documented "save prepacked weights" flow) to memory-map the packed weights. On x86-64 with AVX2
(no AVX-512) and onnxruntime 1.31.1, graphs containing MatMulNBits (int4, block-wise, accuracy_level=4)
compute wrong values after reload. In our autoregressive TTS model the stop token is never emitted, so
generation runs to the 250-frame cap on every input.

Observations (same model, same inputs, Linux x86-64 GitHub runner, CPU flags avx avx2 f16c fma):

• Same on windows-latest x86-64 (AMD Family 25, onnxruntime 1.31.1).
• arm64 (macOS, Linux aarch64) with the same flow is correct; audio is bit-identical to the unpacked model.
• Under Rosetta (no AVX exposed) onnxruntime 1.31.1 does not write int4 prepacked blobs and the issue does not appear.
• A single-node probe (one MatMulNBits node, M=1) compares equal between prepacked and unpacked loads, so the
  mismatch seems to need M>1 or appear only after graph fusions.

To reproduce

Save step (Python):

so = ort.SessionOptions()
so.graph_optimization_level = ort.GraphOptimizationLevel.ORT_ENABLE_ALL
so.intra_op_num_threads = 1
so.add_session_config_entry("session.save_external_prepacked_constant_initializers", "1")
so.add_session_config_entry("session.optimized_model_external_initializers_file_name", "model.packed.data")
so.add_session_config_entry("session.optimized_model_external_initializers_min_size_in_bytes", "1024")
so.optimized_model_filepath = "model.packed.onnx"
ort.InferenceSession("model.onnx", so, providers=["CPUExecutionProvider"])

Then load model.packed.onnx in a new session and compare its outputs with model.onnx. A self-contained
model is not attached yet (see note above); the failing graphs are a text encoder and an autoregressive
prefix step quantized with MatMulNBits (bits=4, block_size=32, symmetric, accuracy_level=4).

Urgency

Workaround in place: we skip prepacking for graphs with MatMulNBits on x86-64, at the cost of about
100 MB more private memory per process.

Platform / versions

Linux (ubuntu-latest) and Windows (windows-latest), x86-64 AVX2; onnxruntime 1.31.1 from PyPI, Python 3.12, CPUExecutionProvider.