Make binary32 numbers interpretable in place

Currently all number encodings except for natural numbers use IEEE 754 binary32 for the longest (4-byte) case. Since there are 30 bits of data the lowest two bits of the mantissa is assumed to be zeroes, but the actual encoding has ones so they can't be read as is. This is particularly annoying when the implementation language doesn't natively support the binary32-uint32 reinterpretation. Since otherwise the choice of those two bits doesn't matter, I propose to flip them.