safe-math: nvc/nvcc/clang still contract multiply-adds, and NvidiaCompiler drops all optimisation #3077

safe-math was defined (#1438, 2020) as "drop -ffast-math" (GCC: -fno-unsafe-math-optimizations; nvc:
drop -fast). Two consequences today:

1. Contraction is unspecified. GCC C99 never contracts; GCC C++ did until dsl: Print double-precision FD weights to 20 digits #3074 (-ffp-contract=off);
   clang defaults to -ffp-contract=on; nvc++ fuses under -O2/-fast; nvcc defaults to --fmad=true.
   test_gradient_equivalence (fp64, safe-math) fails on OpenACC with -fast or -O2 (ratio 1.23) and on
   CUDA by default (1.24), and passes with -Mnofma / --fmad=false. On an RTX 3090 fp64 FMA is worth
   13-14% (OpenACC) and 31-39% (CUDA) of throughput, fp32 nothing, so a blanket "no contraction" is not free
   on GPU.
2. NvidiaCompiler under safe-math appends nothing in place of -fast, so OpenACC safe-math builds are
   effectively unoptimised: 0.3-0.5 GPts/s vs 9-15 with -fast on the same kernels (3D acoustic, 3090).
   Question for maintainers: should safe-math promise "no contraction" on every backend (at the GPU fp64
   cost), or should contraction be a separate, documented per-compiler setting that tests needing exact
   cancellation opt into? Either way NvidiaCompiler should keep a safe optimisation level.