JIT: compile kernels with -fno-math-errno, so gcc can merge repeated sqrt/pow/exp calls #834

Problem

The JIT compiles its pointwise kernels at -O3 -g0 with the platform's compiler. gcc defaults to -fmath-errno: sqrt, pow and exp may set errno, so gcc treats each call as having a side effect. It cannot merge repeated calls with the same argument, and it cannot inline sqrt without a fallback call. An expanded constitutive law repeats the same call many times, once in each place a named quantity appears. Apple clang never sets errno (the Darwin maths library does not), so kernels built on macOS do not pay this cost and kernels built on Linux do. Our kernels never read errno.

Measurement

Linux, the environment's conda-forge gcc 14.3, 2× Xeon Gold 6240R, on a machine loaded by other jobs (the times are noisy; the ratios repeat). The -fmath-errno effect is the cost of the law's pointwise callbacks per quadrature point in one Jacobian assembly, measured by swapping in a constant viscosity on the same mesh and state (scripts/sessions/jit_graph/assembly_split.py on feature/jit-graph-codegen):

With and without the flag, the assembled residual and Jacobian at a fixed state, the solutions, the residual histories and the nonlinear and linear iteration counts are bit-identical, on both fixtures. The generated C does not change.

Proposal

Always pass -fno-math-errno, alongside -std=c99, so that it is kept when UW3_JIT_CFLAGS replaces the optimisation flags. The flags are written into the generated setup.py, which is part of the JIT cache key, so every cached module recompiles once after the change.

Found while measuring the JIT graph route (#823, tier 2).