Make cube non-finite formatting match macOS snprintf

[Copilot]

macOS CI fails because the cube writer hard-codes signed NaN output while macOS snprintf("%13.5E") formats negative NaN differently. The formatter test compares against the platform implementation.

• Non-finite cube values

  • Route NaN and infinity through snprintf("%13.5E").
  • Preserve the writer’s fixed 13-byte field layout.

• Fallback handling

  • Centralize conversion from snprintf output to a 13-byte cube field.
  • Pad safely if formatting returns fewer characters or fails.

if( std::isnan(v) or std::isinf(v) ) {
  format_e13_5_snprintf( v, out );
  return;
}

[wavefunction91]

Heads up on a bug in the plan for this PR: "The formatter test compares against the platform implementation" makes the regression test tautological. If the cube writer's non-finite branch calls snprintf("%13.5E", v) directly, and the test computes its "expected" value the same way (by calling snprintf on the test host), then the test is just checking that the code calls snprintf — it can never fail, even if the writer is broken, because both sides of the comparison are generated by the identical platform-dependent call at run time. That's exactly how the original bug (hard-coded/inconsistent signed-NaN formatting differing from snprintf output) slipped through CI in the first place.

To make this test meaningful, the expected value needs to be a fixed, pre-recorded golden string (captured once from a real macOS snprintf run) rather than a value computed via the same snprintf call inside the test itself. Otherwise this PR "fixes" the CI failure only by removing the test's ability to detect the very platform-formatting mismatch it's meant to catch.