Ensure the substeps we pass to SplitExplicitFreeSurface are the actual used substeps

[simone-silvestri]

in main this is not the case because the number of substeps is adapted to the averaging function by assuming that we cut the interval 2Δt in a substep number of substeps.

This PR calculates the correct ending fractional time based on the averaging function. Typically this corresponds to the time at which the averaging function reaches zero typically between 1 and 2 (times the baroclinic timesteps).
Other kernels might have a more complicated ending time (for example a constant function never reaches zero) for which users have to specify it themselves. For this purpose, this PR adds an AveragingKernel that stores also the ending fractional time and can be passed to the SplitExplicitFreeSurface constructor.

Finally the weights calculation cuts last_fractional_time * Δt in a substep number of substeps.

[glwagner]

I tried to read to the code to figure out what's going on but I didn't get it quickly.

Isn't another option to change the meaning of what the user passes to SplitExplicitFreeSurface? It's unfortunate that these changes make the source code more complicated. Ideally it would simplify the source code.

[glwagner]

Can a fix be to take the user input for substeps and multiply it by 2 before just reusing the existing machinery?