ONNX GraphSurgeon cannot export NumPy-backed INT4 or UINT4 constants #4912

Bug

ONNX GraphSurgeon cannot export a NumPy-backed gs.Constant with ml_dtypes.int4 or ml_dtypes.uint4. constant_to_onnx_tensor packs FLOAT4E2M1 values into ONNX nibbles but writes INT4/UINT4 as one byte per element. ONNX rejects the raw tensor size.

Reproduction

With current main 16ed0636, ONNX 1.19.0 and NumPy 2.2.6:

import ml_dtypes, numpy as np, onnx_graphsurgeon as gs
from onnx_graphsurgeon.exporters.onnx_exporter import OnnxExporter
x = gs.Constant("x", np.array([1, 2, 3, 4], dtype=ml_dtypes.int4))
OnnxExporter.export_tensor_proto(x)

This raises ValueError: Raw data size does not match tensor's size. Expected 2 bytes, but got 4 bytes. UINT4 fails similarly. The exporter should pack each pair of 4-bit values, including odd element counts and signed negative values, and produce an ONNX tensor that decodes to the original values.