ENH: Use NumberToString for lossless float metadata in NRRD IO

_dump_metadata_to_stream<double> and <float> used default stream precision
(6 significant digits), silently truncating user-defined float/double metadata
stored in NRRD key/value fields.  Add explicit template specializations for
double and float that call itk::ConvertNumberToString() for shortest-round-trip
serialization.  Array<double> and Array<float> were already handled losslessly
via the specialized operator<< in itkArrayOutputSpecialization.cxx.

See: #3249

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

Author: hjmjohnson

Modules/IO/NRRD/src/itkNrrdImageIO.cxx

@@ -23,8 +23,10 @@
 #include "itkIOCommon.h"
 #include "itkFloatingPointExceptions.h"
 #include "itkNumericLocale.h"
+#include "itkNumberToString.h"
 
 #include <sstream>
+#include <type_traits>
 
 namespace
 {
@@ -1053,7 +1055,14 @@ _dump_metadata_to_stream(MetaDataDictionary & thisDic, const std::string & key,
   T value;
   if (ExposeMetaData<T>(thisDic, key, value))
   {
-    buffer << value;
+    if constexpr (std::is_same_v<T, float> || std::is_same_v<T, double>)
+    {
+      buffer << ConvertNumberToString(value);
+    }
+    else
+    {
+      buffer << value;
+    }
     return true;
   }
   return false;