Subtraction and comparison against floats is not robust against float underflow

```julia
julia> using FixedPointDecimals

julia> fd = reinterpret(FixedDecimal{Int,0}, (2 << 53) + 1)
FixedDecimal{Int64,0}(18014398509481985)

julia> f = Float64(2 << 53) 
1.8014398509481984e16

julia> isless(f, fd) # this is basically asking 18014398509481984 < 18014398509481985 -> true
false

julia> fd - f # this is basically 18014398509481985 - 18014398509481984 -> 1
0.0
```