test_gradient_equivalence tolerance is tuned, not derived, and only holds without FMA #3078

The fp64 case asserts allclose(grad_u, grad_v, rtol=atol=1e-13) per point. The tolerance was tightened
from 1e-12 to 1e-13 and safe_math=True added on the day #1519 merged (Dec 2020) to make it pass under GCC
C99. It passes because, without contraction, the second time difference of a consistently rounded sequence
cancels under summation by parts; with FMA (g++ before #3074, nvc++, nvcc, clang) the per-point error roughly
doubles and the test sits at or above its tolerance (main passed the C++ jobs with 2% headroom; 13 of 17
small perturbations fail). The natural error scale eps*sum|terms| is up to 2.5e-11 per point, and doubling the
source amplitude alone pushes even the no-FMA case over 1e-13. The norm-relative difference is stable at
1-2e-15 in every configuration.
Proposal: assert ||grad_u - grad_v|| <= 1e-14 ||grad_v|| (and the same for grad_uv) together with a
relaxed per-point check (so a single wrong halo/boundary point still fails), and consider keeping 1/dt**k
outside the time-derivative sum in make_derivative, which makes the test FMA-robust (ratio 0.3 with FMA)
and slightly more accurate, at the cost of touching every time-stepping kernel and the op-count tests.