[caffe2] do not declare __assert_fail in clang builds (#33893)
Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/33893
It appears that when Clang drives CUDA compilation ` __assert_fail` is always defined as device function.
Test Plan:
```lang=bash
buck build mode/opt -c fbcode.cuda_use_clang=true -c cxx.untracked_headers=ignore //fblearner/flow/projects/dper:workflow
```
Reviewed By: ngimel
Differential Revision: D20145034
fbshipit-source-id: 23153411ed631e05421c7afcf41b7ea5619cdd96
