由于 Mat::create() 中的未初始化填充内存,推理结果不是确定的

After investigation, it was found that the problem may be related to the uninitialized memory after Mat::create() calls fastMalloc() to allocate memory. In the case of x86 SIMD Packing (elempack=4/8/16), when the number of channels is not divisible by the pack size, some lanes retain uninitialized data. Subsequent convolutional calculations read these uninitialized values, resulting in differences in the output results for the same input in different runs.