Fix atomic ops for Apple clang/libc++ compatibility

[alisterburt]

LLM spiel:

• Apple's libc++ doesn't support std::atomic<float/double>::fetch_add (C++20 floating-point atomics are incomplete)
• Replace std::atomic-based implementation with OpenMP atomic pragmas, which compile to equivalent hardware instructions.
• Add type trait to deduce real type from complex types, eliminating need for explicit template arguments at call sites.

@dtegunov I ran into issues building this on my work laptop - claude doesn't think there are perf implications to this but a review before merge would be good