Ensure internal ITP bounds are in proper order

[dcourteville]

This is a fix for SciML/DifferentialEquations.jl#1087.

Checklist

• Appropriate tests were added: This can be tested with the Minimal Reproducible Example in the original issue, but the test can't be implemented in DiffEqBase since it needs to actually solve the ODE
• Any code changes were done in a way that does not break public API
• All documentation related to code changes were updated: Nothing to do
• The new code follows the
  contributor guidelines, in particular the SciML Style Guide and
  COLPRAC.
• Any new documentation only uses public API: Nothing to do

Additional context

Internal ITP implementation can sometime return a solution out of tspan when left > right, so we just ensure that left < right even if prob.tspan bounds are inverted when integrating backward. I got the change from the original ITP implementation in NonlinearSolve.jl

[ChrisRackauckas]

Looks right to me. But we should add a test so this doesn't regress.

[dcourteville]

I managed to write a simple test case that reproduce the error, it should be good now.

[oscardssmith]

thanks for fixing this!