MultiLevelPoisson solver tolerance not independent of domain size

[marinlauber]

I just realized (while looking at pressure forces) that the pressure field oscillates when the domain is large (384,256), see attached script.

I have managed to track the issue to the MultiLevelPoisson/solver! function.
This is linked to the V-cycle criterion that checks if the residual on the upper level (r₂) is below a certain tolerance (tol) before performing the Vcycle

while r₂>tol && nᵖ<itmx
      Vcycle!(ml)
      smooth!(p); r₂ = L₂(p)
      log && push!(res,r₂)
      nᵖ+=1
end

The issue is that (r₂) depends on the grid size through p.r as L₂(p::Poisson) = p.r ⋅ p.r, while tol does not. This means that in some cases, the solver! doesn't perform any V-cycle, which doesn't change the pressure field.

The proposed solution is to scale the tol by the grid size, compute it once and store in in the MultiLevelPoisson structure, see #97
TwoD_circle_pressure_check.zip

[weymouth]

Oops! This is what we do in the Fortran code. Not sure how we got away without doing it in WaterLily...