- The inherited operator= is weird, but this seems to work.
- Use TNL::pow instead of std::pow etc.

The splitting with if( std::is_same<...> ) is not perfect, the code for
all devices was always visible to the compiler, and irrelevant parts
were optimized out at a much later stage. To fix compilation for MIC and
custom data type, we need proper template specializations.

- Cleaned up unnecessary specializations of TNL::max and TNL::min
- Implemented result type promotion to fix compilation with nvcc
- Added TNL::sqrt, which will be necessary for MIC

It used to work everywhere, but the compilers on CentOS (on penny)
are weird...

(This version is implemented in Object)

The problem is that the __CUDA_ARCH__ macro is defined only in device
code, so it can't be used for configuring the kernel launches.

This is particularly useful for debugging long simulations which don't
fail soon.