profiler failure on multi-gpu (works on single GPU

[panyx0718]

reproduced in PR: #9051

this line:
"PADDLE_ENFORCE(cudaEventRecord(event_, stream));"

terminate called after throwing an instance of 'paddle::platform::EnforceNotMet'
what(): invalid resource handle at [/paddle/dev/my/Paddle3/paddle/fluid/platform/profiler.cc:73]
PaddlePaddle Call Stacks:
0 0x7f21a215138cp paddle::platform::EnforceNotMet::EnforceNotMet(std::__exception_ptr::exception_ptr, char const*, int) + 572
1 0x7f21a220e66dp paddle::platform::Event::Event(paddle::platform::EventKind, std::__cxx11::basic_string<char, std::char_traits, std::allocator >, unsigned int, paddle::platform::DeviceContext const*) + 317
2 0x7f21a220f316p paddle::platform::PopEvent(std::__cxx11::basic_string<char, std::char_traits, std::allocator > const&, paddle::platform::DeviceContext const*) + 246
3 0x7f21a220f943p paddle::platform::RecordEvent::~RecordEvent() + 211