Minimize IR for atomic float tests (#917)

Removes excessive kernel code from the test examples. The key
source-level change for re-generating such IR from Intel SYCL is
to use `SYCL_EXTERNAL` function definitions instead of "true"
heterogeneous code. E.g. the following DPC++ code may be used to
generate IR for the AtomicFAdd test:

```
using namespace cl::sycl;

template <typename T>
using atomic_ref_T = ONEAPI::atomic_ref<T, ONEAPI::memory_order::relaxed,
                     ONEAPI::memory_scope::device,
                     access::address_space::global_space>;

SYCL_EXTERNAL decltype(auto) AtomicFloatInc(float &Arg) {
  atomic_ref_T<float> Atm(Arg);
  return Atm.fetch_add(1);
}

SYCL_EXTERNAL decltype(auto) AtomicDoubleInc(double &Arg) {
  atomic_ref_T<double> Atm(Arg);
  return Atm.fetch_add(1);
}

Author: Artem Gindinson