Support atomic min/max

[ronawho]

Our atomic support is largely based on C11/C++11, so which operations we support is also largely the same. The main current exception is that we also support operations on real's, but C++20 added support for that too. I think there'd be value in adding support for an atomic min and max as well. We have a few tests that manually implement atomic max (see 9689096677884fcf35b56ba7c22ec5172c65039b) and we have support for min/max reductions, which could be optimized by doing the cross task reductions with atomics.

For other motivation, there's hardware support in some HPC networks and most communication libraries support them (including gasnet, ugni and libfabric.) Rust also supports atomic min/max and there's a proposal to add support in C++: https://www.open-std.org/jtc1/sc22/wg21/docs/papers/2021/p0493r3.pdf.

It's not obvious to me if we'd want to implement min/max as read-conditionally-modify operations or a read-modify-write operations. The C++ proposal above uses RMW to match other operators, and I'd default to following that and optionally doing a RCM if users opt-in to relaxed memory order or something. I think in practice min/max won't be highly contended operations so there probably won't be much of a performance difference (RMW means all writers need exclusive cache line access)