Don't run optimization if expansion is already good to `eps`

For example
```julia
julia> @btime besselj_zero_asymptotic(1, 501)
1.522 ns (0 allocations: 0 bytes)
1574.723079474818

julia> @btime besselj_zero(1, 501)
135.576 ns (0 allocations: 0 bytes)
1574.723079474818

julia> besselj_zero(1, 501) === besselj_zero_asymptotic(1, 501)
true
```

So, `besselj_zero` is 100x slower, but gives the same result bit for bit (for `Float64`) as the asymptotic one.
One could identify the cutoff. Or something more elaborate.