atomic_fetch_add: explicit casts needed for '1' increments

[ndellingwood]

Required when compiling with serial-only backed to avoid compilation
errors due to unmatched types.

decltype used to determine the type based on the input to
atomic_fetch_add; std::remove_reference needed because expressions may
return a reference type.

[mndevec]

@ndellingwood
I though this was going to be fixed in kokkos side. Has that decision changed?

[ndellingwood]

@mndevec the explicit casting for '1's in atomic_fetch_add is necessary after all following modification to the serial implementation for atomics in Kokkos.

I used decltype in this PR to get the type to remove possible errors in guesswork of what the type should be.

The functionality of this PR needed for the kokkos and kokkos-kernels promotion.

[ndellingwood]

@mndevec I had thought this was going to be changed in kokkos also, but seems not after all. The performance fixes are in the kokkos develop branch for serial atomics though.

[ndellingwood]

@mndevec I tested this PR locally with the Serial-only backend enabled, unit tests build and pass.