[PARKED] FASTFLOAT_ASSUME_ROUNDS_TO_NEAREST: compile-time bypass of the rounds_to_nearest() probe

Opt-in macro folding rounds_to_nearest() to true, eliding the volatile-FCMP probe
on the Clinger fast path (parity with ffc's FFC_ROUNDS_TO_NEAREST / EXP-030).
Correctness: bit-identical to default over 200k values in round-to-nearest env
(default build #ifdef-guarded => unchanged). Benchmark (4 Intel gens, c++17+20):
marginal — mesh ~+2.6%, canada ~+1.5%, random neutral/noisy (random uses
Eisel-Lemire, not the Clinger fast path, so no probe to elide). Per-cell noise
incl. small regressions. PARKED: real but below the clean accept bar.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

Author: fcostaoliveira

include/fast_float/parse_number.h

@@ -76,6 +76,13 @@ from_chars_result_t<UC>
  * Credit : @mwalcott3
  */
 fastfloat_really_inline bool rounds_to_nearest() noexcept {
+#ifdef FASTFLOAT_ASSUME_ROUNDS_TO_NEAREST
+  // The caller asserts at compile time that the floating-point environment uses
+  // the default IEEE-754 round-to-nearest mode (i.e. nothing ever changes the
+  // rounding mode via fesetround()). This elides the volatile-float FCMP probe
+  // below, removing it from the Clinger fast path.
+  return true;
+#endif
   // https://lemire.me/blog/2020/06/26/gcc-not-nearest/
 #if (FLT_EVAL_METHOD != 1) && (FLT_EVAL_METHOD != 0)
   return false;