New tensile client

[leekillough]

1. Fix CMakeLists.txt build error
2. Temporarily workaround complex type not being supported (needs to be fixed before release), by converting beta to real before calling ContractionProblem
3. Run clang-format on many files; remove clang-format off in some places (clang-format off was scattered throughout the old Tensile client code, which made it non-uniform from the rest of rocBLAS)
4. Use auto placeholder for the return types of many functions
5. Simplify exception handling try/catch blocks, such as merging all of them in runContractionProblem into one function try/catch block.
6. Remove personal comments containing wbgilmartin name
7. Remove rocblas_create_host_handle and rocblas-bench --lib argument, and change rocblas_create_handle to do the functionality with an environment variable; new Tensile client code must work without changing existing rocBLAS user code, including rocblas-test and rocblas-bench
8. Change mapping from rocBLAS types to Tensile types to use variable templates (C++14) instead of function templates
9. Move tensile_host.hpp from library/include to library/src/include: The public headers in library/include have to be compilable from a C compiler and not use C++ unless guarded with __cplusplus; we do not want to expose C++ internal implementation headers in the public library/include headers
10. Add override default destructor to TensileHostImpl to ensure base class has a virtual destructor, because we use dynamic_cast
11. Simplify some of the path-handling code

Notes:

Need to decide what environment variable name to use for the host path, and what the default path should be if the environment variable is not set.

alpha and beta may be on either the host or the device, depending on handle->pointer_mode. We cannot dereference *alpha or *beta on the host if they are located on device. Ideally we should test the *beta == 0 or *alpha == 0 early exits in the device code rather than in the host code if device pointer mode is in effect; optionally, we may use hipMemcpy to copy from device to host, but that is synchronizing, and not ideal long-term.

Logging needs to work like the original code did; logging occurs whenever handle != nullptr and handle->layer_mode has certain bitmasks set; another PR ROCm#722 is being worked on to change how logging handles nullptr alpha and beta.

The transpose arguments can be rocblas_operation_none, rocblas_operation_transpose or rocblas_operation_conjugate_transpose. For real types, rocblas_operation_transpose and rocblas_operation_conjugate_transpose must operate exactly the same, so normally to detect transpose, we test whether transA != rocblas_operation_transpose_none, rather comparing against rocblas_operation_transpose or rocblas_operation_conjugate_transpose, i.e., if the transpose is not "none", then it represents a transpose regardless of whether it's conjugated.

For complex types, conjugate transpose and regular transpose must be differentiated in certain places. Right now complex types cause linker errors, and std::real(*beta) is used instead of *beta to allow a function overload to work. Complex needs to be added to the overloads, and beta must not be dereferenced if handle->pointer_mode is device pointer mode.

I will be out the rest of the weekend, reading email on my phone occasionally.

@wbgilmartin @zaliu @sdquiring @amcamd @bensander