Reenable FMA in CUDA builds

This might be controversial, but we're kneecapping the CUDA build by
turning off FMA. I think it's high time to turn it back on. Also removes
an inaccurate comment, as FMA doesn't have anything to do with fast
math, as FMA is well specified in IEEE 754-2008.

cmake/detray-compiler-options-cuda.cmake

@@ -29,9 +29,6 @@ if("${CMAKE_CUDA_COMPILER_ID}" MATCHES "NVIDIA")
     # Allow to use functions in device code that are constexpr, even if they are
     # not marked with __device__.
     detray_add_flag( CMAKE_CUDA_FLAGS "--expt-relaxed-constexpr" )
-
-    # Turn off fast math for the device code.
-    detray_add_flag( CMAKE_CUDA_FLAGS "-fmad=false" )
 endif()
 
 # Make CUDA generate debug symbols for the device code as well in a debug