Do not reset default stream for StreamSafeCUDAAllocator

[From00]

PR types

Bug fixes

PR changes

Others

Describe

The default stream in StreamSafeCUDAAllocator will be set when the CUDADeviceContext created.

CUDADeviceContext::CUDADeviceContext(CUDAPlace place) : phi::GPUContext(place) {
  phi::GPUContext::PartialInitWithoutAllocator();
  cuda_stream_.reset(new stream::CUDAStream(phi::GPUContext::stream(), place));
  auto& instance = memory::allocation::AllocatorFacade::Instance();
  instance.SetDefaultStream(place, phi::GPUContext::stream());
  workspace_.reset(new phi::DnnWorkspaceHandle(
      instance.GetAllocator(place).get(), stream()));
}

Normally, the DeviceContextPool is a global singleton and one Place only correspond to one DeviceContext. However, to support multi-stream scheduling, standalone executor creates two extra DeviceContextPools for H2D and D2H stream in StreamAnalyzer, which make one Place correspond to multiple DeviceContext and unexpectedly reset the default stream in runtime.

class StreamAnalyzer {
 public:
  explicit StreamAnalyzer(const platform::Place& place)
      : place_(place), d2h_ctx_pool_({place}), h2d_ctx_pool_({place}) {}
  
  ...

  platform::Place place_;
  platform::DeviceContextPool d2h_ctx_pool_;
  platform::DeviceContextPool h2d_ctx_pool_;
  std::map<size_t, std::shared_ptr<platform::DeviceEvent>> var_id2event_;
};

To avoid this behavior, this PR disables changing default stream after initially setting.

[paddle-bot-old]

你的PR提交成功，感谢你对开源项目的贡献!
请关注后续CI自动化测试结果，详情请参考Paddle-CI手册。
Your PR has been submitted. Thanks for your contribution!
Please wait for the result of CI firstly. See Paddle CI Manual for details.

[zhiqiu]

LGTM