test: cover negative fixed float interpolation

[antono3]

Summary

Retain exact negative f32 and f64 fixed-format interpolation regressions. The implementation fix (signbit(x) != 0) already landed on master in e814006; this refreshed PR is test-only.

Current-master verification (2026-09-26)

Rebased onto 6c359c228241b44f9d9a942696e83b513931e532. The focused regression passes with strict V3 (-new-compiler). It also passes on unmodified master; this PR preserves regression coverage. Formatting verification for all changed V files and git diff --check pass. The previously failing twelve formatter files pass on this master. The unrelated anon_struct_cross_module_init_test.v still fails on unmodified master (the genuinely anonymous parameter case); this refresh does not change it.

[medvednikov]

Reviewed current head d6070ecfec9a4c36d31c4eb1c92febde70a372c2; no concrete correctness issue found in the two added assertions.

Both f32 and f64 are explicitly exercised, and exact fixed-format string comparisons check the negative sign as well as the three decimal places. The chosen values are not halfway rounding cases at that precision, so the assertions isolate sign preservation rather than introducing an ambiguous rounding expectation. This head is test-only; the existing half-away-from-zero cases are retained.

Static test review only; I did not execute the test here, and CI results were not used.