The unpack builtins are implemented using floating-point division, which
is implemented by reciprocal+multiply. As it so happens, 1.0/255.0 falls
almost exactly between two representable 32-bit floats and leads to bad
rounding. None of the other divisors (127, 32767, 65535) have this
property.

Paul was uncomfortable with allowing inexact results in only
unpackUnorm4x8, so allow the results of the builtins (except
unpackHalf2x16) to be within some small epsilon (0.00001) of the actual
value. Some values like 0.0 and 1.0 should be calculated exactly, so
handle them specially.