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

[jlapeyre]

For example

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.