cmp: more accurate handling of large and small numbers

[ear7h]

• comparisons with zero are absoloute (epsilon method)
• comparisons with non-zeros are relative (bit pattern distance method)

I also renamed TOLERANCE to conform with Go conventions

performance difference (slower as expected)

name        old time/op    new time/op    delta
CmpFloat-4    4.89ms ± 1%    5.53ms ± 1%  +12.93%  (p=0.000 n=8+7)

name        old alloc/op   new alloc/op   delta
CmpFloat-4     66.7B ± 1%     75.1B ± 1%  +12.63%  (p=0.001 n=7+7)

name        old allocs/op  new allocs/op  delta
CmpFloat-4      0.00           0.00          ~     (all equal)