UWTN: update the time-derivatives note for the composing solvers

[lmoresi]

AdvDiffusion and NavierStokes now take their transport from the history
manager rather than owning it, and default to EulerianSUPG; the semi-Lagrangian
solvers keep their SLCN names. Verified against development after #688
merged — EulerianSUPG is present with time_derivative, advection,
stabilisation_flux and spatial_weights; AdvDiffusionSUPG correctly does
not exist.

The note

• the DDt hierarchy gains a fifth flavour, and the two Eulerian ones are
  distinguished: one corrects the history explicitly and pays a CFL condition,
  the other assembles the advection in the residual and does not
• "What the Solver Sees" is rewritten in the composing form, keeping the older
  bdf()/adams_moulton_flux() pair as what a semi-Lagrangian manager reduces
  to — the clearest way to show the manager is the plug
• "Choosing a Time Derivative" gets the new defaults and the order= that a
  supplied manager now requires
• ## History entry at 1.1.0, saying what did not change (the scheme theory;
  the viscoelastic stress history, still semi-Lagrangian)

A correction I had to make to my own draft

I first wrote that the default moved because the benchmarks gave "the same
answers, about a tenth of the cost per step". Running the example against the
merged code contradicted both halves:

Courant 0.5, 24 cells	L2 error	s/step
Eulerian SUPG	0.0038	0.023
Semi-Lagrangian	0.0007	0.259
Eulerian	0.1813	0.079

Semi-Lagrangian is more accurate at every Courant number tried, and the design
note's own table puts the cost ratio at 1 : 6.3 rather than a tenth.

The two tests disagree because they measure different things: the design note
carries a blob once around a rotating flow, where interpolation error
accumulates over a circuit and lands SLCN at 21%. The example translates a blob
uniformly, which is the easiest case a semi-Lagrangian scheme ever gets.

So the note now says the default moved on cost and stability, not accuracy,
and says plainly that semi-Lagrangian remains more accurate on smooth
translation — which is why it keeps its place.

The example

examples/timestepping.py — one Gaussian blob, three managers, same solver,
scored against the exact answer. Dimensional (m, s, K) with the
non-dimensionalisation stated where it is applied, per #52. It compares at equal
physical distance rather than equal step count, since a larger Courant number
buying fewer steps is the point of an unconditionally stable scheme, and it says
in its own docstring what it does not show.

Underworld development team with AI support from Claude Code