benchmark测试，开启vulkan出现越界访问的错误

编译环境: win10 vs2019 gtx1660s
复现步骤:
1.下载20220910最新代码，编译
2.执行benchmark.exe 1 16 0 0 1

错误定位：当执行squeezenet_int8模型时，convolution_vulkan.cpp第216-222行出错
此时的weight_data的elemsize为1，但直接用float*进行访问

解决参考：
在Net::load_param函数中（load_param_bin同理），应在
int lr = layer->load_param(pd);
之后执行
if (layer->support_int8_storage)
{
// no int8 gpu support yet
opt.use_vulkan_compute = false;
}
以及
#if NCNN_VULKAN
if (opt.use_vulkan_compute)
layer->vkdev = d->vkdev;
#endif // NCNN_VULKAN

Reproduce the issue on the stated Windows/VS2019 GTX1660S setup with benchmark.exe 1 16 0 0 1, then inspect convolution_vulkan.cpp lines 216-222 and the Net::load_param and load_param_bin entry points. Verify the squeezenet_int8 path and Vulkan handling around int8 storage. Done means the benchmark no longer performs the reported out-of-bounds access.