docs: correct LBL W convention, LaTeX typo, NH3 note

[olemke]

Summary

Three documentation corrections in the line-by-line (LBL) docs, each bringing the
text in line with what the code actually does. No functional or API changes.

Changes

1. Relaxation matrix index convention — doc/arts/dev.lbl.rst, doc/arts/user.lbl.rst

The prepared real relaxation matrix is stored as W[to, from], the coupling
into line to from line from, following the untransposition of the old
Fortran copy. The docs still described it as W[from, to]. Corrected in both the
developer array-preparation table and the lbl.relaxation_matrix_profile() user
description.

This now agrees with the implementation and the nanobind docstring:

• src/core/lbl/lbl_lineshape_voigt_ecs_hadded.cpp assigns W[to, from]
• relaxation_matrix_profile documents W[i, j] as "the coupling into line i
  from line j", with the optical sum rule as the row closure
  sum_j dipole[j] * W[i, j] == 0

2. LaTeX typo — doc/arts/concept.absorption.lbl.rst

\Delta_nu_Z -> \Delta\nu_{Z} in the complex Voigt argument z, so the Zeeman
shift renders as a subscripted symbol instead of literal text.

3. NH3 example output claim — doc/arts/user.lbl.rst, tests/core/lbl/ecs_nh3.py

The docs stated the NH3 example "saves ecs_nh3.png in the current directory".
It never does: the script only calls plt.show(), which ARTS_HEADLESS
suppresses under CTest, and no test in tests/core/lbl writes a figure file. The
documentation is corrected rather than adding a savefig call, matching the
behaviour of the other ECS tests.