Problems compiling with the NVHPC suite

[pratikvn]

The NVHPC suite is a SDK from NVIDIA, that combines all the high performance compilers and libraries and provides it to the computing clusters. Most clusters have this available as a module now.

The host compiler is an adaptation of the PGI compiler and NVHPC additionally includes the nvcc compiler and the necessary CUDA libraries (Cusparse, Cublas etc,)

Currently Ginkgo fails with NVHPC due to two main reasons:

1. NVHPC does not seem to support user-defined reductions, which we need due to templating and given that we need to define reductions for our own types.
2. NVHPC seems to have some issues with Rapidjson, with some typedefs in the biginteger.h header, which the host compiler does not seems to recognize.

Fixed issues: We previously also faced issues with the unified kernel reduction but that was fixed in PR #926.

[upsj]

1. should be solved when #861 is finished and merged :)

[tcojean]

Should be fixed after #1331. Please reopen if that is not the case.