Try Rodas5P as default solver

[visr]

It runs Dommel at similar speed to QNDF, and doesn't diverge over long timespans when iterating between drainage and infiltration on the bucket test model.

HWS model is 44s with Rodas5P vs 10s QNDF. Though QNDF is known to be wrong (#1851). With QNDF I get one large leak of -632063 and several 1e-8. With Rodas5P just one of 2.5e-6. So this addresses that as well.

On top of #1868.

TODO: fix tests, look into backwater model performance

[visr]

Rodas5P doesn't seem to like ManningResistance very well. The Julia tests go from 15 min to 30 min, and fail and are slow at Manning tests:

  test/run_models_test.jl                          |  137     5    142  18m28.1s
    trivial model                                  |   32           32     12.6s
    bucket model                                   |    8            8      0.5s
    leaky bucket model                             |   14     2     16     21.1s
    basic model                                    |   12           12      0.4s
    basic arrow model                              |    3            3      1.6s
    basic transient model                          |    5     1      6      0.3s
    Allocation example model                       |    3            3     20.3s
    sparse and AD/FDM jac solver options           |    9            9     55.4s
    TabulatedRatingCurve model                     |    5            5     23.7s
    Profile                                        |   12           12      0.1s
    Outlet constraints                             |    3            3     21.3s
    UserDemand                                     |    5            5     14.2s
    ManningResistance                              |    3     2      5  15m25.0s
    mean_flow                                      |   23           23     11.8s

[visr]

Closing in favor of #1874.