fix interpolation for GPUVern7 and GPUVern9

[jfeist]

The interpolation for GPUVern7 and GPUVern9 used the current time t after the last timestep as the "reference" time to evaluate the function, but should use the previous time tprev (so that the function is evaluated within the interval [tprev,t]). This gave wrong interpolated values when using saveat for any time-dependent ODE. A very simple example (simply solving $u'(t) = p t$, i.e., $u(t) = p t^2/2$):

using OrdinaryDiffEqCore
using DiffEqGPU
using DiffEqGPU: vectorized_asolve, make_prob_compatible
using StaticArrays
using JLArrays

f(u, p, t) = SA[p * t] 
ode_prob(p, tmax) = ODEProblem{false}(f, SA[0.0], (0.0, tmax), p)

t = 0:0.1:20
ps = 0:0.1:1.
sq_ana = ps' .* t.^2 ./ 2

probs = ode_prob.(ps, t[end])
jlprobs = JLArray(make_prob_compatible.(probs))

for alg in (GPUTsit5, GPUVern7, GPUVern9)
    ts, us = vectorized_asolve(jlprobs, probs[1], alg(); saveat=t, reltol=1e-12, abstol=1e-12)
    sq_num = first.(us)
    println("alg: ", alg, "  err: ", sum(abs2,sq_num .- sq_ana))
end

On master, I get:

alg: GPUTsit5  err: 5.0895151945504214e-23
alg: GPUVern7  err: 76368.43220161786
alg: GPUVern9  err: 42300.10632221898

After the fix, it is

alg: GPUTsit5  err: 5.0895151945504214e-23
alg: GPUVern7  err: 4.705062089752212e-21
alg: GPUVern9  err: 5.6572466674974e-19

[ChrisRackauckas]

Yes thanks!