Improve float conversions in PI controllers

[ChrisRackauckas]

This is fairly hard to test, but it basically stems from float(1//50)::Float64, which means that this was always doing some float64 stuff, even when it should've been doing fastpow stuff. The test case just really required a Float32:

using DiffEqCallbacks, OrdinaryDiffEq, Tracker

Base.prevfloat(r::Tracker.TrackedReal) = Tracker.track(prevfloat, r)
Tracker.@grad function prevfloat(r::Real)
    prevfloat(Tracker.data(r)), Δ -> (Δ,)
end
Base.nextfloat(r::Tracker.TrackedReal) = Tracker.track(nextfloat, r)
Tracker.@grad function nextfloat(r::Real)
    nextfloat(Tracker.data(r)), Δ -> (Δ,)
end

function rober(u, p::TrackedArray, t)
    y₁, y₂, y₃ = u
    k₁, k₂, k₃ = p
    return Tracker.collect([-k₁ * y₁ + k₃ * y₂ * y₃,
        k₁ * y₁ - k₂ * y₂^2 - k₃ * y₂ * y₃,
        k₂ * y₂^2])
end

p = TrackedArray([1.9f0, 1.0f0, 3.0f0])
u0 = TrackedArray([1.0f0, 0.0f0, 0.0f0])
tspan = TrackedArray([0.0f0, 1.0f0])
prob = ODEProblem{false}(rober, u0, tspan, p)

p = TrackedArray([1.9f0, 1.0f0, 3.0f0])
u0 = TrackedArray([1.0f0, 0.0f0, 0.0f0])
tspan = TrackedArray([0.0f0, 1.0f0])
prob = ODEProblem{false}(rober, u0, tspan, p)
saved_values = SavedValues(eltype(tspan), eltype(p))
cb = SavingCallback((u, t, integrator) -> integrator.EEst * integrator.dt, saved_values)

solve(remake(prob, u0 = u0, p = p, tspan = tspan),
            Tsit5(),
            sensealg = SensitivityADPassThrough(),
            callback = cb)
        
@test !all(iszero.(Tracker.gradient(
    p -> begin
        solve(remake(prob, u0 = u0, p = p, tspan = tspan),
            Tsit5(),
            sensealg = SensitivityADPassThrough(),
            callback = cb)
        return sum(saved_values.saveval)
    end,
    p)[1]))

Thus downstream tests with FastPower.jl used catches this, and it's somewhat hard to construct a case that's this sensitive to the type.