fix initdt calculation

[oscardssmith]

fixes #2492.
That said, @ChrisRackauckas our initdt selection does seem like we never considered the case where iszero(u0). When this happens, our d₀ is 0, so our dt₀ = smalldt, which puts a hard cap on our return value of 100*smalldt, which is often incredibly pessimistic. Would it be reasonable to have the 2nd stage of this algorithm loop when 100dt₀ < dt₁ (i.e. we will do an extra function call at t + dt₁ to see if our dt₁ guess was reasonable. This would cost an extra function eval, but in this case, our dt₁ is 300x bigger than our d₀, which will require taking multiple steps with basically no progress before our timestepping is reasonable.

[ChrisRackauckas]

Would it be reasonable to have the 2nd stage of this algorithm loop when 100dt₀ < dt₁ (i.e. we will do an extra function call at t + dt₁ to see if our dt₁ guess was reasonable. This would cost an extra function eval, but in this case, our dt₁ is 300x bigger than our d₀, which will require taking multiple steps with basically no progress before our timestepping is reasonable

Just doing some starting acceleration might make more sense. But that doesn't sound so unreasonable.