[flang][runtime] Fix CUDA flang-rt build breakage (#135220)

I used "std::nullopt" instead of the correct "Fortran::common::nullopt"
in a recent patch, and you can get away with that only for CPU builds.
Fix.

flang-rt/include/flang-rt/runtime/io-stmt.h

@@ -156,7 +156,7 @@ public:
       if (at_ && at_ < limit_) {
         return *at_;
       } else {
-        return std::nullopt;
+        return Fortran::common::nullopt;
       }
     }
     RT_API_ATTRS void NextRecord(IoStatementState &io) {
@@ -200,7 +200,7 @@ public:
         byteCount = ch ? 1 : 0;
         return ch;
       } else if (!field->MustUseSlowPath()) {
-        return std::nullopt;
+        return Fortran::common::nullopt;
       }
     }
     return GetCurrentCharSlow(byteCount);