Created by: luotao1

To make developers build Paddle much faster, this PR changes the default value of CUDA_ARCH_NAME from All to Auto.

• Speedup benchmark data in #5713: The GPU compiling time: about (14min ~ 16min) -> about 9min in TeamCity
• Print warning when use Auto:

-- CUDA detected: 8.0
-- WARNING: This is just a warning for publishing release.
      You are building GPU version without supporting different architectures.
      So the wheel package may fail on other GPU architectures.
      You can add -DCUDA_ARCH_NAME=All in cmake command
      to get a full wheel package to resolve this warning.
      While, this version will still work on local GPU architecture.
-- Added CUDA NVCC flags for: sm_35

• fix typo on architectures