Fix CUDA int8x4 vectorize

[llehtahw]

• fix CodeGenCUDA::PrintVecElemStore and CodeGenCUDA::PrintVecElemLoad
  • avoid wrong generated cuda code like this:

    int _x; // int as int8x4, from #1569
    _x.x = x; // or x = _x.x

  • get the deleted test(Use int for int8x4 due to performance overhead of char4 #1569) back
    • the test passed but cannot benefit from vectorize
• fix accumulation of shared/local memory usage with vector types
  • test updated

@vinx13 @nishi-t @tqchen

[tqchen]

@vinx13 please manage this PR

[vinx13]

Thanks @llehtahw this is merged