Fix autodiff performance

[visr]

We currently default to finite differencing of the Jacobian due to it being much faster, see #1832. Autodiff generally should be more accurate and a bit faster, so perhaps there is some place we need to provide a rule. A profile shows not much difference with autodiff on or off, but there are many more timesteps. Possibly that means that the derivative is wrong, we should check it against finite diff.

[visr]

While working on #1835 I saw something strange. The backwater model tests started failing with finite diff. The actual heads were just over a meter higher on the upstream end, but were similar on the downstream end. I set backwater to use autodiff for now, but we should look into this later. It could also be some interaction with standard_step_method.

ManningResistance: Test Failed at /home/runner/work/Ribasim/Ribasim/core/test/run_models_test.jl:509
  Expression: all(isapprox.(h_expected, h_actual; atol = 0.02))