Implement 'float' and 'double' constants

[anachrocomputer]

Use IEEE-754 32-bit and 64-bit floating-point formats. Maybe even make ANSI 'long double' work for 128-bit F.P. But no, I don't fancy writing an F.P. library.