Use true node spacing for finite-volume node-to-edge shifts

[rtimms]

Fixes #5754.

The bug

FiniteVolume.shift's node to edge matrices hard-code weights that assume
uniformly spaced nodes:

• arithmetic_mean interpolates interior edges with [0.5, 0.5]
• both arithmetic_mean and harmonic_mean extrapolate the two exterior edges
  with [1.5, -0.5] and [-0.5, 1.5]

boundary_value_or_flux extrapolates using the true node positions instead. On a
uniform mesh the two agree and nothing is wrong. On a non-uniform mesh they do
not, and that breaks mass conservation wherever a flux is imposed as a gradient
divided by a surface value — which is exactly what the Fickian particle does
(fickian_diffusion.py:209):

f"{Domain} {phase_name}particle bc [mol.m-4]": -j * R_nondim / self.param.F / pybamm.surf(D_eff)

The flux that condition produces is -D_eff * grad(c), and D_eff reaches the
surface edge through node_to_edge (harmonic, since the operand is a gradient).
When that value disagrees with surf(D_eff), the imposed surface flux is off by
their ratio — a systematic bias in the same direction on every step, so it
accumulates monotonically rather than averaging out.

With a concentration-dependent particle diffusivity and an
Exponential1DSubMesh, ORegan2022 gains 26 % of its particle lithium over
three 1C cycles. Nothing reports a loss or gain; Total lithium in particles [mol] just drifts. In a long ageing run with LAM and SEI active it looks like a
degradation signal, and LLI [%] stops reconciling with the model's own loss
terms.

The fix

Use the real node spacing for both:

• interior edges are interpolated with distance weights
  (d_edges[i+1] / (d_edges[i] + d_edges[i+1]) on node i), matching the
  reasoning already used for harmonic_mean's beta
• the two exterior edges are extrapolated from the true node positions, shared
  by both means via a small exterior_edge_rows helper

Both reduce to the old hard-coded values on a uniform mesh, so uniform-mesh
results are bit-for-bit unchanged.

Result

Drop in Total lithium in particles [mol] over three 1C cycles, stock parameter
sets, SPMe (negative = lithium created):

parameter set	r_n	uniform, before/after	exponential, before	exponential, after
ORegan2022	16	-0.0000	-25.9351	-0.0000
ORegan2022	64	0.0000	-0.3540	-0.0000
Ecker2015	16	0.0000	0.4274	0.0000
Ecker2015	64	-0.0000	0.0244	0.0000
Chen2020 (constant D)	16	0.0000	0.0000	0.0000

Tests

Adds test_node_to_edge_exterior_matches_boundary_value, parametrised over both
means. On an Exponential1DSubMesh with a linear field it asserts that the
exterior edge values equal boundary_value and equal the exact field value
there, and that arithmetic reproduces the linear field at every edge. Against
the old weights this gives 1.0100 and 3.0013 where the exact values are 1 and 3.

Full unit suite passes. test_idaklu_jax.py has three pre-existing failures on
this machine (ProcessedVariable indexing a 0-d time array); they reproduce
identically on unmodified main.

Note on scope

The exterior-edge rows are what break conservation. The interior [0.5, 0.5]
weights are a separate first-order accuracy error on non-uniform meshes; I fixed
both since they share a cause, but I'm happy to split the interior change out if
you'd rather keep the PR to the conservation bug.

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 98.24%. Comparing base (6e5baab) to head (0d77486).
⚠️ Report is 1 commits behind head on main.

Additional details and impacted files

@@            Coverage Diff             @@
##             main    #5755      +/-   ##
==========================================
+ Coverage   98.15%   98.24%   +0.08%     
==========================================
  Files         341      341              
  Lines       33710    33715       +5     
==========================================
+ Hits        33088    33123      +35     
+ Misses        622      592      -30     

☔ View full report in Codecov by Harness.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.