fix(generation): spin-polarised inputs start every moment at zero

[junwen94]

What is emitted

generation.py:224-227 is the whole of the magnetic input:

if advice.spin_orbit.enabled:
    lines.extend(["  noncolin = .true.", "  lspinorb = .true."])
elif advice.magnetism.spin_polarized:
    lines.append("  nspin = 2")

starting_magnetization appears nowhere in the repository, and neither do
angle1 or angle2.

Why that is a problem

The collinear branch starts every species at zero moment. Quantum ESPRESSO
defaults starting_magnetization to 0. With nspin = 2 and no starting
moments the two spin channels are identical, so the SCF has no symmetry to
break and normally settles on the non-magnetic solution. The input says
spin-polarised and the calculation is not.

This is the quiet direction of the error. The run converges, reports a total
magnetisation of zero, and nothing about it looks wrong.

The spin-orbit branch discards the magnetism advice entirely. The elif
means that whenever an operator passes spin_orbit_coupling=True, whatever
_advise_magnetism concluded is dropped. Skipping nspin = 2 is correct --
it is not allowed alongside noncolin -- but the advice is not translated into
the non-collinear form either. A non-collinear calculation needs
starting_magnetization together with angle1 and angle2 per species to
define the directions; with none of them, the moments are zero and the
directions undefined.

So a magnetic heavy-element system, which is the case SOC exists for, gets the
most expensive spin treatment available and no magnetism at all.

This is a contracts gap, not only a generation one

MagnetismAdvice carries spin_polarized, magnetic_elements, and
provenance. There is no field for a moment, an ordering, or a direction, so
generation.py has nothing it could write even if it wanted to. Fixing this
means deciding what magnetism advice contains before deciding how to print it.

What is needed

1. A moment on MagnetismAdvice, per species or per site.
2. Emit starting_magnetization whenever nspin = 2 is emitted. Even a
   crude non-zero default is strictly better than zero, because zero is the one
   value guaranteed not to produce the calculation the input claims.
3. Translate magnetism into the non-collinear branch rather than dropping
   it, which needs angle1/angle2 and therefore a direction in the advice.
4. Decide what an ordering means here. An antiferromagnetic arrangement may
   not fit the chemical cell, and converging one can require a magnetic
   supercell -- which changes the structure every later recommendation is made
   for, k-mesh included. That makes magnetism the first setting whose answer can
   feed backwards through the pipeline, and it should be settled deliberately
   rather than discovered later.

Item 2 is worth doing on its own, ahead of the rest.

Not in scope

The relativistic pseudopotential linkage is already correct:
_advise_pseudopotentials sets relativistic_mode = "full" when SOC is
enabled, and warns when heavy elements are present without it.

Related: #175 records the heuristic that decides spin_polarized in the first
place. goldilocks-ml's magnetism scaffolding records the three quantities a
model would have to supply -- is_magnetic, ordering, and
magnetic_moments -- and notes that only the third buys a magnetic
calculation.