Created by: luotao1

In general, it's recommended to use PADDLE_ENFORCE for error_mesage. However, in CUDA device kernel or HOSTDEVICE function, we must use PADDLE_ASSERT now. https://github.com/PaddlePaddle/Paddle/blob/c2063217e7458cfa97a16874b6a6982448259304/paddle/fluid/platform/assert.h#L40 This PR：

• replace part of PADDLE_ASSERT to PADDLE_ENFORCE, and later we will see how to unify PADDLE_ASSERT and PADDLE_ENFORCE, which is so confused for usage now.
• remove unused fallback_alloc_size_ in system_allocator.cc and add unit-test for CUDAPinnedAllocator, to pass the PR_CI_COVERAGE