server异步response写pb消息耗时较长

**Describe the bug (描述bug)**
brpc server 异步响应模式在往response中写消息字段时，通过前后取时间戳发现，response写pb消息偶现耗时4-5ms，更长时甚至20ms+

response proto格式：
{
optional uint64 id = 1;
optional int32 stat = 2;
optional uint32 cxxdiali = 3;
repeated uint32 acount = 4; //大概40+个
}

在response赋值pb消息前后取时间戳
// send response
**long long tik = GetClock();**
response->set_id (request->id ());
if (succ) {
response->set_stat (STATUS_OK);
} else {
response->set_stat (STATUS_FAILURE);
}
response->set_cxxdiali (carty);
for (int i = 1; i < size; ++i) {
response->add_ acount(xxdialis[i]);
}
**long long tok = GetClock();**

当tok-tik 超时3ms以上记录，占比大概2~3/1000：
cost: 5 1628476500710->1628476500715.
cost: 6 1628476503023->1628476503029.
cost: 8 1628476503323->1628476503331.
cost: 4 1628476505146->1628476505150.
cost: 4 1628476506299->1628476506303.
cost: 5 1628476507680->1628476507685.
cost: 23 1628476508124->1628476508147.
cost: 24 1628476508274->1628476508298.
cost: 23 1628476508455->1628476508478.
cost: 23 1628476508515->1628476508538.
cost: 23 1628476508545->1628476508568.
cost: 23 1628476508755->1628476508778.
cost: 23 1628476508876->1628476508899.
cost: 23 1628476508936->1628476508959.
cost: 23 1628476509056->1628476509079.

**To Reproduce (复现方法)**

**Expected behavior (期望行为)**

**Versions (各种版本)**
OS: linux 3.10.0-514.44.5.10.h254.x86_64
Compiler： gcc 9.x
brpc: 0.9.7
protobuf: 3.6.1

**Additional context/screenshots (更多上下文/截图)**

我的服务占用内存很高，我判断是pb在写消息时申请内存导致的偶现延时增大，请问各位是否有什么优化建议？
我查了pb的资料有arena方式创建message，不知道brpc内是否支持，如何调用？或者各位大佬有其他好的建议？

多谢 ！

Start by reproducing the asynchronous brpc server response path described in the issue, using the listed brpc 0.9.7 and protobuf 3.6.1 versions. Measure the timestamped response field assignments and repeated-field additions under the reported workload; done means identifying the source of the latency spikes and validating an optimization or documented explanation.

Written by the indexing model from the issue text.