Keep the digit-separator feature off the default hot path

PR #369's "performance-neutral" claim did not hold: the default
(no-separator) parse was ~70% larger in the hot frame (parse_number_string
inlined into from_chars: 833 -> 1414 x86/arm64 instructions), because

  1. parse_number_string is force-inlined (always_inline), and the runtime
     dispatch inlined BOTH the has_separator==true and ==false instantiations
     into the default caller -- dragging the entire separator-aware scanner
     into the hot frame; and
  2. the runtime separator branch survived in from_chars() even though a
     default-constructed parse_options provably has no separator, because the
     thin from_chars_caller forwarder was not inlined, so the compile-time
     '\0' was lost across the call boundary and the branch (plus two calls)
     could not be folded away.

Fixes:
  * FASTFLOAT_NOINLINE + a cold parse_number_string_with_separator trampoline
    so the separator instantiation stays strictly out of line; the default
    caller only ever materializes the has_separator==false code.
  * Force-inline from_chars_caller::call (all three specializations) so the
    separator branch constant-folds away on the common from_chars() path.

Net: the default hot frame returns to 844 instructions (vs 833 on main; the
+11 is the unrelated parse_number_string refactor and is unchanged by this
commit). The separator path is unaffected and all tests pass.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

include/fast_float/float_common.h

@@ -217,6 +217,12 @@ using parse_options = parse_options_t<char>;
 #define fastfloat_really_inline inline __attribute__((always_inline))
 #endif
 
+#ifdef FASTFLOAT_VISUAL_STUDIO
+#define FASTFLOAT_NOINLINE __declspec(noinline)
+#else
+#define FASTFLOAT_NOINLINE __attribute__((noinline, cold))
+#endif
+
 // Branch-probability hint marking the rare slow-path branches as cold, so the
 // optimizer keeps the out-of-line slow-path re-parse off the hot path (and does
 // not duplicate the force-inlined hot scanner into the caller, which bloated

include/fast_float/parse_number.h

@@ -139,7 +139,7 @@ fastfloat_really_inline bool rounds_to_nearest() noexcept {
 
 template <typename T> struct from_chars_caller {
   template <typename UC>
-  FASTFLOAT_CONSTEXPR20 static from_chars_result_t<UC>
+  fastfloat_really_inline FASTFLOAT_CONSTEXPR20 static from_chars_result_t<UC>
   call(UC const *first, UC const *last, T &value,
        parse_options_t<UC> options) noexcept {
     return from_chars_advanced(first, last, value, options);
@@ -149,7 +149,7 @@ template <typename T> struct from_chars_caller {
 #ifdef __STDCPP_FLOAT32_T__
 template <> struct from_chars_caller<std::float32_t> {
   template <typename UC>
-  FASTFLOAT_CONSTEXPR20 static from_chars_result_t<UC>
+  fastfloat_really_inline FASTFLOAT_CONSTEXPR20 static from_chars_result_t<UC>
   call(UC const *first, UC const *last, std::float32_t &value,
        parse_options_t<UC> options) noexcept {
     // if std::float32_t is defined, and we are in C++23 mode; macro set for
@@ -166,7 +166,7 @@ template <> struct from_chars_caller<std::float32_t> {
 #ifdef __STDCPP_FLOAT64_T__
 template <> struct from_chars_caller<std::float64_t> {
   template <typename UC>
-  FASTFLOAT_CONSTEXPR20 static from_chars_result_t<UC>
+  fastfloat_really_inline FASTFLOAT_CONSTEXPR20 static from_chars_result_t<UC>
   call(UC const *first, UC const *last, std::float64_t &value,
        parse_options_t<UC> options) noexcept {
     // if std::float64_t is defined, and we are in C++23 mode; macro set for
@@ -289,23 +289,32 @@ from_chars_advanced(parsed_number_string_t<UC> &pns, T &value) noexcept {
   return answer;
 }
 
+template <bool bjf, typename UC>
+FASTFLOAT_NOINLINE FASTFLOAT_CONSTEXPR20 parsed_number_string_t<UC>
+parse_number_string_with_separator(UC const *first, UC const *last,
+                                   parse_options_t<UC> options,
+                                   bool store_spans) noexcept {
+  return parse_number_string<bjf, true, UC>(first, last, options, store_spans);
+}
+
 // Runtime -> compile-time dispatch over both boolean knobs of
 // parse_number_string. basic_json_fmt was already dispatched this way; the
 // digit separator is selected here too so that the separator-aware code paths
-// stay confined to the (cold) has_separator==true instantiation. Callers that
-// never set a separator -- the overwhelming majority -- run the
+// stay confined to the (cold, out-of-line) has_separator==true instantiation.
+// Callers that never set a separator -- the overwhelming majority -- run the
 // has_separator==false instantiation, which is byte-for-byte the original
-// separator-free parser.
+// separator-free parser; the separator check is a single predictable branch
+// into cold code.
 template <typename UC>
 fastfloat_really_inline FASTFLOAT_CONSTEXPR20 parsed_number_string_t<UC>
 parse_number_string_options(UC const *first, UC const *last,
                             parse_options_t<UC> options, bool bjf,
                             bool store_spans) noexcept {
-  if (options.digit_separator != UC('\0')) {
-    return bjf ? parse_number_string<true, true, UC>(first, last, options,
-                                                     store_spans)
-               : parse_number_string<false, true, UC>(first, last, options,
-                                                      store_spans);
+  if fastfloat_unlikely (options.digit_separator != UC('\0')) {
+    return bjf ? parse_number_string_with_separator<true, UC>(
+                     first, last, options, store_spans)
+               : parse_number_string_with_separator<false, UC>(
+                     first, last, options, store_spans);
   }
   return bjf ? parse_number_string<true, false, UC>(first, last, options,
                                                     store_spans)