Remove SafeSinh cross-lane defense now that eve fixed it upstream

[foolnotion]

Summary

#172 defended eve::sinh/eve::tanh call sites against a confirmed
third-party bug: a NaN in one SIMD lane of an eve::wide register
corrupted the computed result in a completely unrelated, non-NaN lane
of the same register (affecting sinh, tanh, expm1, coth, csch,
gd in eve's math module).

jfalcou/eve#2366 fixes the
shared root cause: expm1's own non-finite-lane handling was calling
the fast/approximate [raw] path for the entire register whenever any
lane was non-finite, degrading accuracy in every other lane too. Since
sinh, tanh, and coth call expm1 directly, and csch/gd call
sinh/tanh respectively, this one-line upstream fix covers all 6
affected functions.

Changes

• Bumped the foolnotion/nur-pkg (eve package) pin to
  unstable-2026-08-07 (past eve#2366's merge).
• Removed SafeSinh (functions.hpp) and reverted its 3 call sites
  (primal Sinh, JIT Sinh, Cosh's derivative rule) to call
  eve::sinh directly.
• Left Tanh's derivative rule on FastTanh — Defend Sinh/Tanh-derivative call sites against eve's cross-lane NaN bug #172 framed that switch
  as also fixing an unrelated inconsistency with primal Tanh (which
  already used FastTanh), not purely a defensive workaround, so kept
  it independent of the eve bug status.
• Updated the stale SafeSinh-referencing comment on the regression
  test, kept the test itself as a permanent guard.

Verification

• Standalone before/after repro (built at operon's actual
  -march=x86-64-v3 target) against both eve pins: all 6 functions
  (sinh, tanh, expm1, coth, csch, gd) show cross-lane
  corruption under the old pin, clean under the new one.
• Full test suite (~[performance]) passes both with the defense in
  place (confirms the eve bump alone doesn't regress anything) and
  without it (confirms nothing else depended on SafeSinh
  incidentally) — 0 failures in both runs.

Test plan

• Standalone cross-lane repro, old vs new eve pin
• Full suite passes with defense removed