rdar://104608391

Currently, DFG will attempt to speculate Int32 for certain double constants during the fixup phase. If a number constant is an Int32, or falls within a valid range (between ±2^48), it's allowed to be truncated into an integer. However, NaN values cannot be compared and ordered, so unlike other non-finite values like positive or negative Infinity, they can pass this range check and be converted into an Int32 despite clearly being non-integral values. We should add a check for NaN constants and reject Int32 speculation if we come across one during the fixup phase.