Issue on GPU when calling kernels outside DPLASMA

[QingleiCao]

Describe the bug

When calling kernels from DPLASMA on GPU, either setup_parsec/cleanup_parsec or parsec_info_register/parsec_info_unregister need to be called in users' codes. Otherwise, DPLASMA::CUDA::HANDLES will not be registered.

It would be better to hide this on the users' side, so they can directly call APIs from DPLASMA.

To Reproduce

Steps to reproduce the behavior:

1. Checkout version 'commit-hash'
2. Compile with the following options '....' [e.g., head config.log]
3. Run test '....' [e.g. ctest --output-on-failure -R xyz]
4. See error

Expected behavior

A clear and concise description of what you expected to happen.

Environment (please complete the following information):

• DPLASMA version: [e.g. git hash]
• PaRSEC status: external/internal and version: [e.g., git hash]
• OS: [e.g. CentOS/7]
• Compiler: [e.g. GCC/7.3.0]
• MPI version: [e.g. Open MPI/4.1.2]

Additional context

Add any other context about the problem here.
The content of the config.log file can be useful in some cases.