Atomic function allow implicit conversion of update argument

[crtrott]

Most backends allow a call like this to happen:

long* ptr = ...
atomic_fetch_add(ptr,2);

This is enabled by an explicit available overload of:

void atomic_fetch_add(long* ptr, const long val);

For some backends (like Serial only) we only have this function:

template<class T>
void atomic_fetch_add(T* ptr, const T val);

This will fail to match since the template deduction is ambiguous, it could either be long or int.
To make this match we could do the following:

template<class T>
void atomic_fetch_add(T* ptr, typename std::add_const<T>::type& val);

Now only the first parameter is used for the actual deduction of T and val follows matching rules which include implicit conversions.

[crtrott]

I am fixing one of those cases for Serial, since I can't get FENL to compile otherwise.

[ibaned]

@crtrott I think you managed to fix all of these?