Implement GPU-aware IO

[ia267]

This PR implements GPU aware IO (closes #176) where data is written directly from device memory to ADIOS2 without host staging.

Process:

• After defining the variable, tell ADIOS2 the buffer lives on GPU
• Pass the device pointer via C wrapper (the Fortran API call adios2_put does not accept CUDA device array. Therefore a C wrapper adios2_put_gpu is created to receive the device pointer and calls the C adios2_put. Without the C wrapper it triggers a device-to-copy which defeats the purpose)
• At runtime, write-field subroutines detects the cuda backend and uses the GPU path to write. Non-cuda paths uses the default host path

In order to use GPU-aware IO we have to use the compiler flag: -DWITH_ADIOS2_GPU_AWARE=ON

As part of this PR I have also implemented single-precision writing from GPU, as well as reading mixed-precision (closes #276) and as well fixed a bug in checkpoints, where it was using the host pointer (%data) see 20d98bf.