Add infrastructure for compile-time CUB tests

[gevtushenko]

Description

closes #1123

This PR introduces necessary infrastructure for compile-time CUB tests. Tests that ends with _fail are treated as compile-time tests. Example:

// %PARAM% TEST_ERR err 0:1

int main()
{
  #if TEST_ERR == 0
  static_assert(false, "fail one"); 
  #elif TEST_ERR == 1
  static_assert(false, "fail two"); // expected-error-1 {{"fail two"}}
  #endif
}

Test above is compiled twice, for TEST_ERR = 0 and TEST_ERR = 1. In the first case, ctest relies on compiler return code to detect issue. In the second case, ctest expects fail two message in the compiler output. More usage examples can be found in the following tests:

• cub/test/test_device_radix_sort_decomposer_fail.cu
• cub/test/test_fail.cu
• cub/test/test_param_fail.cu
• cub/test/test_param_return_code_fail.cu
• cub/test/test_return_code_fail.cu.

Checklist

• New or existing tests cover these changes.
• The documentation is up to date with these changes.

[alliepiper]

LGTM -- this is very useful!